Thinking in patterns:
representations in the neural basis of theory of mind
by
Jorie Koster-Hale
B.A. Linguistics and Cognitive Science
Pomona College, 2009
Submitted to the Department of Brain and Cognitive Sciences
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Neuroscience
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
September 2014
© 2014 Jorie Koster-Hale. All rights reserved.
The author hereby grants MIT permission to reproduce and to distribute publicly paper
and electronic copies of this thesis document in whole or in part in any medium now known or hereafter
created.
Signature of Author:______________________________________________________
Jorie Koster-Hale
Department of Brain and Cognitive Sciences
August 16, 2014
Certified by: __________________________________________________________
Rebecca R. Saxe
Professor of Brain and Cognitive Sciences
Thesis Supervisor
Accepted by:__________________________________________________________
Matthew A. Wilson
Sherman Fairchild Professor of Neuroscience
Director of Graduate Education for Brain and Cognitive Science
1 of 179
This page intentionally left blank
2 of 179
Thinking in patterns:
representations in the neural basis of theory of mind
by
Jorie Koster-Hale
Submitted to the Department of Brain and Cognitive Sciences
on August 6, 2014, in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Neuroscience
Abstract
Social life depends on understanding other people’s behavior: why they do the things they
do, and what they are likely to do next. These actions are just observable consequences of
an unobservable, internal causal structure: the person’s intentions, beliefs, and goals. A
cornerstone of the human capacity for social cognition is the ability to reason about these
invisible causes; having a “theory of mind”. A remarkable body of evidence has
demonstrated that social cognition reliably and selectively recruits a specific group of brain
regions. Yet, we have little insight into how these neural substrates function at a
computational level. This thesis lays the groundwork to address that question, both
empirically and theoretically, first by demonstrating that functional neuroimaging can find
behaviorally relevant features of mental state representation within the cortical regions that
support social cognition, and second by proposing a theoretical framework to interpret
activity in these brain regions. In Chapter 1, I review the literature of the last 15 years, and
argue that a key next step in understanding the neural basis of social cognition is
characterizing the neural representations and computations supported by “social” brain
regions. In Chapter 2, I demonstrate in four experiments that functional neuroimaging can
be used to find neural representations of distinct features of mental states. Specifically, I
show that multivoxel pattern analysis (MVPA) can detect features of mental state
representations (e.g., intent), and that these neural patterns are behaviorally relevant,
including in autism spectrum disorders. In Chapter 3, I demonstrate that these brain regions
contain explicit, abstract representations of another feature of others’ mental states:
perceptual source. I find that these representations persist in the face of drastic changes in
developmental history (congenital blindness), providing evidence that these representations
emerge even in the absence of relevant first-person experience. In Chapter 4, I demonstrate
that these cortical regions contain representations of epistemic and emotional features of
others’ beliefs, and that these features are represented along continuous, abstract
dimensions. Finally, in Chapter 5, I extend a model from vision and neuroeconomics —
predictive coding — and explore its application to the neural basis of social cognition.
Together, this work provides a key next step to understanding the neural basis of theory of
mind, by demonstrating that it is possible to find abstract, behaviorally relevant features of
mental state inferences inside cortical regions that support social cognition, and taking a first
step in characterizing their content and format.
Thesis Supervisor: " Rebecca R. Saxe
"
"
"
Associate Professor of Brain and Cognitive Sciences
3 of 179
This page intentionally left blank
4 of 179
Contents
ACKNOWLEDGEMENTS!
9
CHAPTER 1. FUNCTIONAL NEUROIMAGING OF THEORY OF MIND!
11
Introduction!
11
Theory of Mind and the Brain!
12
Theory of mind brain regions respond to mental state content"
Specificity"
Links to behavior"
Summary"
A Strong Hypothesis!
13
22
24
26
27
Objections from theoretical considerations"
Empirical objections"
Summary"
What Now?!
27
33
35
36
Differences between theory of mind regions"
Information within theory of mind regions"
Patterns within theory of mind regions
Magnitude: “more” theory of mind
A First Step!
36
38
41
CHAPTER 2. DECODING MORAL JUDGMENTS FROM NEURAL REPRESENTATIONS OF
INTENTIONS !
45
Introduction!
46
Methods!
48
Participants"
fMRI Protocol and Task"
Experiments 1 and 4
Experiment 2
Experiment 3
Theory of Mind Localizer task
Acquisition and Preprocessing"
Motion and Artifact Analysis"
fMRI Analysis"
Functional Localizer: Individual subject ROIs
Within-ROI Magnitude Analysis
Within-ROI Pattern Correlation Analysis
Whole Brain Pattern Correlation Analysis (Searchlight)
Results!
48
48
52
52
52
55
fMRI task: Behavioral Results"
Motion and Artifact Analysis Results"
fMRI Results: Functional Localizer"
55
55
55
5 of 179
fMRI Results: Response Magnitude"
fMRI Results: Pattern Analysis for Neurotypical"
Within ROI Pattern Analysis for Experiment 1
Within ROI Pattern Analysis for Experiment 2 and 3
Behavioral and Neural Correlation
Whole Brain Searchlight Analysis for Experiment 1,2,3 (NT)
fMRI Results: Pattern Analysis for ASD"
Within ROI Pattern Analysis for Experiment 4 (ASD)
Behavioral and Neural Correlation (ASD)
Whole Brain Searchlight for Experiment 4 (ASD)
ASD symptom severity in Experiment 4 (ASD)
Group Comparison: NT vs ASD"
Within ROI Pattern Analysis
Behavioral and Neural Correlation
Decoding true vs. false beliefs?"
Discussion!
56
58
61
61
62
63
MVPA discriminates accidental and intentional harms in RTPJ of neurotypical adults"
Pattern discrimination of intentional versus accidental harm is correlated with moral
judgment"
High functioning adults with ASD show atypical neural activity in RTPJ"
Conclusion"
63
64
65
66
Acknowledgments!
67
Appendix!
67
SI Behavioral Results: Results from Raw Behavioral Responses"
Supplementary Results: Pairwise matched subsets of ASD and NT"
Independent effects of magnitude and pattern"
A note on separating the phases of Experiment 1"
67
68
69
70
CHAPTER 3. THINKING ABOUT SEEING: PERCEPTUAL SOURCES OF KNOWLEDGE ARE
ENCODED IN THE THEORY OF MIND BRAIN REGIONS OF SIGHTED AND BLIND ADULTS ! 71
Introduction!
72
Methods!
76
Participants"
Comparison to Bedny et al. (2009)"
fMRI Protocol and Task"
Acquisition and Preprocessing"
Motion and Artifact Analysis"
fMRI Analysis"
Functional Localizer: Individual subject ROIs
Within-ROI Pattern Correlation Analysis
Whole Brain Pattern Correlation Analysis (Searchlight)
Results!
76
76
77
78
79
79
82
fMRI task: Behavioral Results"
Sighted participants
82
6 of 179
Blind participants
Group Comparison
Motion and Artifact Analysis Results"
fMRI Results: Functional Localizer"
fMRI Results: Within ROI Pattern Analysis"
Sighted participants
Blind participants
Across Groups
Whole Brain Searchlight Analysis"
Source (seeing vs. hearing)
Valence (good vs. bad):
Discussion!
83
83
83
86
88
MVPA reveals features of mental state representations"
Knowledge source and theory of mind"
The valence of others’ mental states"
Blind adults represent knowledge source"
Learning about sight"
Conclusion"
Acknowledgments!
88
89
90
91
92
93
94
CHAPTER 4. MENTALIZING REGIONS REPRESENT CONTINUOUS, ABSTRACT DIMENSIONS
OF OTHERS’ BELIEFS !
95
Introduction!
96
Methods!
98
Participants"
Stimuli"
fMRI Task"
Theory of Mind Localizer"
Acquisition and Preprocessing"
Motion and Artifact Analysis"
fMRI Analysis"
Modeling
Whole-brain random effects (univariate)
Defining Individual subject ROIs
Within-ROI Multivariate Analysis (MVPA)
Item-Wise classification: Testing for continuous representations
Results!
98
98
102
102
102
103
103
108
fMRI task: Behavioral Results"
fMRI task: Classification Results"
Quality
Modality
Quality vs Modality
Directness
Valence
Generalization to other manipulations of evidence quality"
7 of 179
108
108
111
Stimuli
Methods and Analyses
Results
Discussion!
114
Evidence Quality is explicitly represented in RTPJ and RMSTS"
The modality of others’ evidence is explicitly represented in right and left TPJ"
No region could decode directness: first-person experience versus hearsay"
The valence of others’ emotions is represented in VMPFC "
Neural representations of abstract features"
Why track epistemic features of beliefs?"
Summary and Conclusion"
Acknowledgments!
114
115
115
116
116
118
119
119
CHAPTER 5. THEORY OF MIND: A NEURAL PREDICTION PROBLEM !
121
Introduction!
122
A predictive coding framework!
123
The sources of predictions!
126
Predicting goal directed action"
Predicting beliefs and desires"
Predicting preferences and personalities"
Beyond error!
128
130
132
133
Distinguishing prediction from attention"
Finding the prediction signal"
Using predictive coding to test the neural computations underlying theory of mind"
Acknowledgements!
133
136
137
138
CHAPTER 6. GENERAL DISCUSSION!
Summary!
139
139
Where We Are!
140
Content in theory of mind brain regions"
A feature-based approach to mental state representations"
Beyond Features"
Continuous, abstract representations
Inputs to representations
Replication and generalization
What next?!
140
141
143
146
Information between regions, and between networks"
Development"
Links to behavior: Content and structure of mental state inference
Beyond neuroimaging"
REFERENCES !
146
147
148
151
8 of 179
Acknowledgements
This thesis came into existence with the help and support of many people.
The first thanks goes to my advisor, Rebecca Saxe. Without her thoughtfulness,
intelligence, generosity, support, and excitement, my work would not be what it is, and I
would be not the scientist that I am.
Many thanks to my committee, Nancy Kanwisher, for her clarity in thinking, impeccable
standards, and enthusiasm; Josh Tenenbaum, for always seeing and pushing for the
bigger picture (and making “prior” part of my non-scientific vocabulary), and Fiery
Cushman, for his sharp insights, excitement for science, and clear sense of how things
fit together. Thanks to Amy Skerry, Stefano Anzellotti, Hilary Richardson, and Martin
Hackl who read through, edited, and thought about many portions of this dissertation,
and to Todd Thompson, who spent many hours writing with me in cafes around
Cambridge.
Many thanks to all the people who made MIT the best place possible to get a PhD.
Thanks to people in Saxelab: Marina, Liane, Emile, Mina, Lindsey, Dorit, Stefano, Hyo,
Amy, Ben, Alex, Hilary, David, Alek, Nick, Julianne, Nir, Adele, Laura, Swetha, Natalia
and Mika. Thanks to BCS folk who made our department fun and interesting, especially
folks in Nancy’s lab, Laura’s lab, and Josh’s lab, and to the imaging team downstairs:
Steve, Sheeba, and Atsushi. Thanks to my non-MIT collaborators, Naomi, Rachel M.,
Rachel B., Jenny, Jennie, and Amy, and thanks to the people who nudged me in the
right direction when I arrived at MIT, especially Irene Heim, Adam Albright, and Jim
DiCarlo.
Thanks in particular to Liane and Marina, who are not only great co-authors, but
amazing role models; to Nir and Nick for answering random Matlab questions at 2am; to
Hilary for being a fabulous officemate, coauthor, and coffee partner; to Amy for always
being there to talk through ideas (scientific and otherwise); to Todd for writing dates,
walks, talks, and The Epic Computer Rescue of 2014, to Aeriel for 27 years of cookie
and coffee date, to Sarah for surprisingly effective mindmelding across hundreds of
miles, and to Natasha for reminding me that the best part of getting a PhD is getting to
wear wizard robes for a day.
Thanks to my family, who not only put up with me writing this dissertation in the middle
of a family reunion, but supported me with a continual stream of baked goods. (On the
note of sustenance, thanks also to Area Four for putting good caffeine within walking
distance, and to Tatte for butter cookies.) Thanks to my dad, William Earl, for always
thinking my work was so cool. Thanks to my moms, Pamela and Mindy, for being the
best parents ever, and thanks to Martin for being Most Excellent.
9 of 179
This page intentionally left blank
10 of 179
Chapter 1.
Functional neuroimaging of theory of mind1
Introduction
Social life depends on developing an understanding of other people’s behavior: why
they do the things they do, and what they are likely to do next. Critically, though, actions
are just observable consequences of an unobservable, internal causal structure: the
person’s goals and intentions, beliefs and desires, preferences and personality traits. A
cornerstone of the human capacity for social cognition is the ability to reason about
these invisible causes: if a person checks her watch, is she uncertain about the time or
bored with the conversation? And is she chronically rude or just unusually frazzled? The
ability to reason about these questions is sometimes called having a “theory of mind
(ToM)”.
In the last decade, the number of papers that use human neuroimaging tools to
investigate the neural basis of theory of mind has exploded from four to, as of 2014, well
over 500. Studying ToM with neuroimaging works. Unlike many aspects of higher-level
cognition, which tend to produce small and highly variable patterns of responses across
individuals and tasks, ToM tasks generally elicit activity in an astonishingly robust and
reliable group of brain regions. In fact, convergence came almost immediately. By 2000,
Frith and Frith (2000) concluded that “studies in which volunteers have to make
inferences about the mental states of others activate a number of brain areas, most
notable the medial [pre]frontal cortex [(mPFC)] and temporo-parietal junction [(TPJ)].”
These regions remain the focus of most neuroimaging studies of ToM and social
cognition, more than a decade later (Adolphs, 2009; 2010; Carrington, 2009; Van
Overwalle & Van Overwalle, 2009). This consensus is one of the most remarkable
scientific contributions of human neuroimaging, and the one least foreshadowed by a
century of animal neuroscience.
Nevertheless, most of the fundamental questions about how our brains allow us to
understand other minds remain unanswered; we have mainly discovered where to look.
This dissertation offers the first step in closing that gap: demonstrating that not only do
the parts of cortex that support social cognition contain abstract, behaviorally relevant
features of mental state inference, but that we can find them using fMRI.
In this chapter, I offer a perspective on the contribution that neuroimaging has made to
the science of ToM in the last decade, some thoughts on the contribution that it could
make in the next one, and lay out the specific contributions of this thesis. The chapter
has four sections. Theory of Mind and the Brain reviews the existing evidence for a
basic association between thinking about people’s thoughts and feelings and activity in
1A
version of this chapter appeared as Koster-Hale, J., & Saxe, R. R. (2013). Functional neuroimaging of
theory of mind. In S. Baron-Cohen, H. Tager-Flusberg, & M. V. Lombardo, Understanding Other Minds
Perspectives from developmental social neuroscience (pp. 132–163). Oxford University Press.
11 of 179
this group of brain regions. A Strong Hypothesis discusses some objections, both
theoretical and empirical, to a strong interpretation of this association. What Next?
highlights newer approaches to functional imaging data, how they can contribute to the
future neuroscience of ToM, and their strengths and limitations. A First Step lays out
the goal of the thesis and provides a roadmap for the subsequent chapters.
Theory of Mind and the Brain
Over the course of development, human children make a remarkable discovery: other
people have minds, both similar to and distinct from their own. Other people see the
world from a different angle, have different desires and preferences, and acquire
different knowledge and beliefs. Children learn that other people’s minds contain
representations of the world which are often true and reasonable, but which may be
strange, incomplete, or even entirely false. These discoveries (i.e. “building a Theory of
Mind”) help children to make sense of some otherwise mystifying behaviors: why mom
would eat broccoli, even though there is chocolate cake available, (e.g. Repacholi &
Gopnik, 1997), or why she is looking for the milk in the fridge, even though dad just put
it on the table (e.g. Wimmer & Perner, 1983).
Developmental psychologists historically focused on one key transition in this
developmental process—when and how children come to understand false beliefs.
Assessing understanding of false beliefs has been taken to be a good measure of ToM
capacity because it requires a child to understand both that someone can maintain a
representation of the world, and that this representation may not match the true state of
reality or the child’s own beliefs. In a standard version of the false belief task, children
might see that while their mother thinks the milk is in the fridge (having put it there 5
minutes ago), it is now actually on the table. The children are asked: “Where will she
look for the milk?” or “Why is she looking in the fridge?”. Five-year-old children, like
adults, usually predict that she will look in the fridge, because that is where she thinks
the milk is (Wellman, Cross, & Watson, 2001). Three-year-olds, however, predict that
she will look on the table, explaining that she wants the milk and the milk is on the table
(Heyes, 2014 for further discussion of ToM behavior in pre-verbal children; at least when
asked explicitly; see e.g. Onishi, Baillargeon, & Baillargeon, 2005; Saxe, 2013;
Southgate, Senju, & Csibra, 2007). In fact, when three-year-olds see her look in the
fridge instead, some will go so far as to fabricate belief-independent explanations,
stating that she no longer wants the milk, and must be looking for something else
(Wellman et al., 2001; Wimmer & Perner, 1983)
Building off decades of experience in developmental psychology, the first neuroimaging
studies of ToM also used versions of false belief tasks. Adults, lying in positron emission
tomography (PET) or magnetic resonance imaging (MRI) scanners, read short stories
describing a person’s action (see Figure 1.1), and were asked to explain that action
(usually silently to themselves, to avoid motion artifacts). These early studies revealed
increased levels of blood oxygen and glucose uptake (indirect measures of metabolic
activity, henceforth called “activity”), in a small, but consistent group of brain regions—
left and right TPJ, anterior regions of the superior temporal sulcus (STS), down to the
temporal poles, mPFC, and medial parietal cortex (precuneus, PC; Figure 1.1).
12 of 179
Activity during the false belief task is, of course, far from sufficient evidence that these
brain regions have any role in understanding other minds. This proposition becomes
more compelling after reviewing the range of different experimental tasks and
paradigms that have been used successfully over the years, across laboratories and
countries, each aiming to elicit some aspect of “understanding other minds.” While
some studies used complex verbal narratives, other researchers have used simple
sentences or non-verbal cartoons; some studies explicitly instruct participants to think
about a person’s thoughts, and others have elicited ToM spontaneously. The
heterogeneity of methods, materials, and participant demographics makes it especially
impressive that these studies have converged on the same regions of the brain. In the
next section, I review a sample of these different procedures.
Theory of mind brain regions respond to mental state content
In the original theory of mind neuroimaging experiments, both the content of the
materials and the explicit instructions focused participants’ attention on (or away from)
thinking about someone’s mind. For example, in an early PET study, Fletcher Happe,
Frith, Baker, Dolan, Frackowiak, et al. (1995) told participants that they would be
reading different kinds of verbal passages. Just before each item, the participant was
told what kind of story was coming next. If it was a “mental” story, participants were
instructed that it was “vital to consider the thoughts and feelings of the characters,” and
then shown a story revolving around someone's mental states (see example in Table
1.1). If the story was a “physical” story, participants were first instructed that thinking
about thoughts and feelings was irrelevant and undesirable, then shown a control story
(see example in Table 1.1). After each story, participants were instructed to silently
answer an action-explanation question, such as “Why did the prisoner say that?”.
Glucose consumption increased in the theory of mind brain regions while people read
the mental stories, relative to the control stories.
The same design has been used with non-verbal stimuli. In an early fMRI experiment
(Gallagher et al., 2000), participants both read the stories used by Fletch et al. (1995)
and were shown cartoons depicting visual jokes that either relied on ToM (in which
understanding the joke depended on attribution of either a false belief or ignorance), or
other types of humor (such as puns, idioms, and physical humor). Again, participants
were cued in advance about whether to expect a mental or control cartoon (for
examples, see Figure 1.2). For both types of cartoons, they were asked to silently
contemplate the meaning; for mental cartoons, they were also explicitly instructed to
consider the thoughts and feelings of the characters. With less than 20 minutes of
scanning for each task, Gallagher and colleagues found that the same group of brain
regions showed increased activity for both verbal and non-verbal ToM stimuli; these
regions include the bilateral temporal-parietal junction and the middle prefrontal cortex.
13 of 179
Figure 1.2: Examples from Gallagher et al. (2000). ToM cartoon, non-ToM cartoon, jumbled picture
This activation profile does not depend on explicit instruction to attend to mental state
information. Some experiments elicit thinking about thoughts simply by describing those
thoughts in words. For example, participants can answer questions about people’s
mental characteristics, such as “How likely is Queen Elizabeth to think that keeping a
diary is important?” vs. their physical traits, such as “How likely is Queen Elizabeth to
sneeze when a cat is nearby?” (Lombardo et al., 2010 Table 1.1); or they can read
single sentences describing thoughts (“He thinks that the nuts are rancid”) or facts (“It is
likely that the nuts are rancid;” Zaitchik et al., 2010). In both cases, the items related to
mental states elicited more activity in ToM regions than the control conditions.
Other experiments elicit thinking about thoughts indirectly. Saxe & Kanwisher (2003)
used verbal stories based on either inferences about false beliefs or about physical
events, similar to Fletcher et al. (1995). Participants were not given any explicit
instructions about the different kinds of stories, with a common task across conditions.
In Experiment 1, participants did not give any response, while in Experiment 2, they
responded to fill-in-the-blank questions about details in the stories. Similarly, Mason &
Just (2011) had participants read short stories about actions, and then answer simple
comprehension questions. Critically, the stories elicited spontaneous inferences about
unstated, but implied, events; some of these inferences were about a character’s
thoughts (mental), and some about purely physical events (control). Compared with the
original Fletcher et al. (1995) stories, the stories in these experiments (see examples in
Table 1.1) were shorter, and included less (or no) explicit description of thoughts and
feelings (see also Bruneau & Saxe, 2012). Instead, the thoughts and feelings of the
characters had to be inferred. Listening to the mental stories elicited strong activation in
ToM regions relative to control stories, suggesting consideration of the character’s
thoughts despite the absence of explicit instruction.
14 of 179
Mental
Control
During'the'war,'the'Red'army'capture'a'member'of'the'
Blue'army.'They'want'him'to'tell'them'where'his'armies''
tanks'are.'They'know'that'they'are'either'by'the'sea'or'in'
the'mountains.'They'know'that'the'prisoner'will'not'
want'to'tell'them,'he'will'want'to'save'his'army,'and'so'
he'will'certainly'lie'to'them.'The'prisoner'is'very'brave'
and'very'clever,'he'will'not'let'them'find'his'tanks.'The'
tanks'are'really'in'the'mountains.'Now'when'the'other'
side'ask'him'where'his'tanks'are'he'says,'"They'are'in'the'
mountains."
Two'enemy'powers'have'been'at'war'for'a'very'long'
time.'Each'army'has'won'several'battles,'but'no'the'
outcome'could'go'either'way.'The'forces'are'equally'
matched.'However,'the'Blue'army'is'stronger'than'the'
Yellow'army'in'air'foot'soldiers'and'artillery.'But'the'
yellow'army'is'stronger'than'the'Blue'army'in'air'power.'
On'the'day'of'the'final'battle,'which'will'decide'the'
outcome'of'the'war,'there'is'heavy'fog'over'the'
mountains'where'the'fighting'is'about'to'occur.'LowG
lying'clouds'hang'over'the'soldiers.'By'the'end'of'the'
day,'the'Blue'army'has'won.
Brad'had'no'money,'but'just'had'to'have'the'beautiful'
ruby'ring'for'his'wife.'Seeing'no'salespeople'around,'he'
quietly'made'his'way'closer'to'the'counter.'He'was'seen'
running'out'the'door.
While'playing'in'the'waves,'Sarah’s'Frisbee'went'flying'
toward'the'rocks'in'the'shallow'water.'While'searching'
for'it,'she'stepped'on'a'piece'of'glass.'Sarah'had'to'wear'
a'bandage'on'her'foot'for'a'week.
The'path'to'the'castle'leads'via'the'lake.'But'children'tell'
the'tourists:'‘‘The'way'to'the'castle'goes'through'the'
woods.’’'The'tourists'now'think'that'the'castle'is'via'the.'
woods/lake?
The'sign'to'the'monastery'points'to'the'path'through'the'
woods.'While'playing'the'children'make'the'sign'point'to'
the'golf'course.According'to'the'sign'the'monastery'is'
now'in'the'direction'of'the.'golf'course/woods?
How'likely'is'Queen'Elizabeth'to'think'that'keeping'a'
diary'is'important?
How'likely'is'Queen'Elizabeth'to'sneeze'when'a'cat'is'
nearby?
John'was'on'a'hike'with'his'girlfriend.'He'had'an'
engagement'ring'in'his'pocket'and'at'a'beautiful'
overlook'he'proposed'marriage.'His'girlfriend'said'that'
she'could'not'marry'him'and'began'crying.'John'sat'on'a'
rock'and'looked'at'the'ring.
Joe'was'playing'soccer'with'his'friends.'He'slid'in'to'steal'
the'ball'away,'but'his'cleat'stuck'in'the'grass'and'he'
rolled'over'his'ankle,'breaking'his'ankle'and'tearing'the'
ligaments.'His'face'was'flushed'as'he'rolled'over.
That'morning,'people'sat'around'looking'at'each'other,'
wondering'if'they'were'dreaming,'because'everything'
looked'purple.'Some'people'were'shocked.'Some'people'
thought'that'it'was'funny'to'see'everybody'all'purple.'
But'even'the'smartest'scientists'didn’t'know'what'had'
happened.'
The'whole'world'had'turned'purple'overnight.'Just'about'
everything'was'purple,'included'the'sky'and'the'ocean'
and'the'mountains'and'the'trees.'The'tallest'skyscrapers'
and'the'tiniest'ants'were'all'purple.'The'bicycles'and'
furniture'and'food'were'purple.'Even'the'candy'was'
purple.'
In'spite'of'her'neighborhood,'Erica'has'a'strong'dislike'of' Erica'lives'in'Los'Angeles.'One'night'recently'she'was'in'a'
violence,'and'believes'that'conflicts'can'usually'be'
bar'where'a'fight'broke'out'between'two'drunk'men'and'
resolved'without'fists.'
she'was'caught'in'between.
Sam'thinks'he'can'grow'trees'with'fruit'that'taste'like'
pizza.'How'likely'is'it'that'Sam'wants'these'trees'for'a'
treehouse'too?
In'the'backyard'are'trees'with'fruit'that'taste'like'pizza'
when'ripe.'How'likely'is'it'that'these'trees'can'be'used'
for'building'a'treehouse?
Table 1.1. Samples of experiments that elicit thinking about thoughts and feelings by
manipulating the content of the stimuli. Sample stimuli from Fletcher et al. 1995, Mason & Just 2010,
Perner et al. 2006, Lombardo et al. 2010, Bruneau et al. in 2013, Saxe et al. 2009, Saxe & Wexler 2005,
Young et al. 2010.
15 of 179
In fact, when participants read a story, they appear to automatically represent the
thoughts and feelings of the characters in order to make sense of the plot, even if
instructed to perform an orthogonal task. For example, Koster-Hale & Saxe (2011) had
participants read short verbal stories, and then make a delayed-match-to-sample
judgment, indicating whether a single probe word occurred in the story (match) or not
(non-match); half of the stories described a false belief and half described physical
representations. Despite the word-level task (and no mention of ToM in the explicit
instructions), the contrast revealed activation across the ToM network. Similarly, in two
fMRI experiments with children aged 5–12 years, children heard child-friendly verbal
stories, describing characters’ thoughts and feelings vs. physical events; children
answered orthogonal (delayed-match-to-sample) questions about each story. As with
adults, there was increased activation in the ToM network when children were listening
to stories involving thoughts and feelings (Gweon, Dodell-Feder, Bedny, & Saxe, 2012;
Saxe, Whitfield-Gabrieli, Scholz, & Pelphrey, 2009).
Sommer, Döhnel, Sodian, Meinhardt, Thoermer, & Hajak (2007) also use non-verbal
stimuli to focus participants’ attention on the thoughts of a character, but without
explicitly cuing the condition. They showed participants a series of cartoon images
depicting a story—Betty hides her ball, Nick moves it, and then Betty comes back to
look for her ball. In half of the trials, Betty looks into the box where she thinks the ball is
(the expected condition); in the other half, she looks into the other box (the unexpected
condition). Participants judged whether, based on the character’s beliefs, the
character’s action was expected or unexpected. Rather than explicitly labeling the
mental and control conditions, a key contrast was between the beginning of the trial,
before mental state inferences were possible, and the end of the trial, when participants
had presumably made belief inferences to complete the task. The second key contrast
was between trials in which Betty had a false belief (so that predicting her action
required considering her thoughts) and trials in which Betty knew where the ball was all
along (so her action could be predicted based on the actual location of the ball).2 Both
contrasts revealed activity in ToM brain regions (Figure 1.3).
As well as depicting complex stories lines, non-verbal stimuli can be endowed with
mentalistic content by altering the movements of simple geometric shapes (Heider &
Simmel, 1944). For example, in a PET study, Castelli, Happé, Frith, & Frith (2000)
showed participants animations of two triangles moving around. Participants were
instructed that while some of the triangles “just move about with random movement [...]
disconnected from each other” (the control condition), other animations would show
“two triangles doing something more complex together, as if they are taking into account
their reciprocal feelings and thoughts [...] for example, courting each other,” (the mental
2
In the original study, this contrast could have been due to a difference between false vs. true beliefs, or
between representing a belief (required for the false trials) and making a prediction based solely the
actual location of the ball (possible for the true beliefs). Subsequent work has shown that activation
observed by Sommer et al. 2007 is due to the latter: individuals often simply reduce the information they
need to process by choosing not to represent true beliefs as a mental state (Apperly et al., 2007), and
when this is controlled for, neuroimaging has revealed indistinguishably high activation for true and false
beliefs (Döhnel, Schuwerk, Meinhardt, Sodian, Hajak, & Sommer, 2012; Jenkins & Mitchell, 2010; Young
et al., 2010b).
16 of 179
condition). Participants watched the animations, and then described what the triangles
were doing. ToM animations elicited more activity than the random animations in the
TPJ, the nearby superior temporal sulcus (STS), and the mPFC.
Figure 1.3: A False Belief trial from Sommer et al. (2007)
Another procedure for eliciting spontaneous ToM in the scanner was developed by
Spiers & Maguire (2007). Participants engaged in naturalistic actions (e.g. driving a
taxicab through bustling London streets) in a rich virtual reality environment (Figure
1.4). After the scan and without prior warning, participants reviewed their performance,
and were asked to recall their spontaneous thoughts during each of the events. These
recollections were coded for content concerning the thoughts and intentions of the
taxicab customers and the other drivers and pedestrians on the road (e.g. “I reckon that
she’s going to change her mind”) and used these coded events to predict neural
responses. They found that when participants were thinking about someone else’s
intentions, but not during other events, regions in the ToM network showed increased
activity.
Other studies have targeted spontaneous consideration of others’ intentions by asking
participants to make moral judgments (see Young & Waytz, 2013 for a review). Morally
relevant facts appear to rely in part on consideration of intentions (Cushman, 2008), and
evoke increased activity in the ToM network, relative to other facts in a story (Young &
Saxe, 2009). The same brain regions are recruited when participants are forced to
17 of 179
choose between acting on a personal desire vs. a conflicting moral principle, compared
to deciding between two conflicting personal desires (Sommer et al., 2010). These
regions are also recruited in children watching an animation of one person intentionally
harming another, compared to animations of other painful and non-painful situations
(Decety, Michalska, & Akitsuki, 2008).
Figure 1.4: Example stimuli from Spiers and Maguire (2006)
18 of 179
Figure 1.5: Examples of cartoons used in Schnell et al. (2010) and Walter et al. (2010). For each
frame, participants were asked to decide whether (a) the mood of the protagonist got better or worse, or
(b) the number of living things they could see increased or decreased.
Thinking about thoughts and feelings can be also manipulated by changing the task
while holding the stimuli constant. In an early PET study, Goel, Grafman, Sadato, &
Hallett (1995) showed participants sets of 75 photographs of objects, some modern and
familiar, and some from pre-fifteenth century North American aboriginal culture.
Participants either judged whether the object was elongated along the principle axis (the
control task) or whether “someone with the background knowledge of Christopher
Columbus could infer the [object’s] function” (the mental task). They found increased
ToM activation when participants were considering Christopher Columbus, but not when
making the physical judgments. Similarly, Walter and colleagues showed participants
sequences of three cartoon images (Schnell, Bluschke, Konradt, & Walter, 2010; Walter
et al., 2010). Participants either judged, on each picture, whether “the protagonist feels
worse/equal/better, compared to the previous picture” (the mental task) or whether “the
number of living beings [in the image is] smaller/equal/greater, compared to the
previous picture” (the control task, Figure 1.5).
female
male
Figure 1.6: Reading the Mind in the Eyes (Baron-Cohen, 1991; Adams et al. 2009)
In another set of studies, Baron-Cohen and others (Adams et al., 2010; BaronCohenMO'Riordan, 1999; Platek, 2004) showed participants pictures of a person’s
eyes. Participants pressed a button to indicate either the mental/emotional state of the
19 of 179
person in the picture (e.g. embarrassed, flirtatious, worried; the mental task) or their
gender (the control task, Figure 1.6). Mitchell, Banaji, & MacRae (2005) asked
participants to either judge either how happy a person was to be photographed (mental)
or how symmetric their face was (control). In all of these cases, the mental tasks
activated the ToM regions more than the control tasks.
Spunt and colleagues have recently developed a clever paradigm for eliciting ToM using
a simple task manipulation. In their first experiment, Spunt, Satpute, & Lieberman
(2010) showed participants pictures of simple human actions, (e.g. a person riding a
bike), and instructed them to silently answer one of three questions: why the person is
doing the action (e.g. to get exercise), what the person is doing (e.g. riding a bike), or
how the person is doing it (e.g. holding handlebars, Figure 1.7). These questions
require successively less consideration of the mind of the person and, correspondingly,
showed successively less ToM region activity. Spunt & Lieberman (2012) replicated the
result using a similar paradigm with brief naturalistic film clips of facial expressions of
emotions. Participants judged either how the person is expressing their emotion (e.g.
“looking down and away,” the control task) or why she is feeling that emotion (e.g. “she
is confused because a friend let her down,” the mental task). Again, ToM regions were
recruited more when thinking about why than how.
Figure 1.7 Spunt et al. (2010): parametric design of why, what and how?
Using these types of parametric designs and analyzing the continuous magnitude of
response in ToM regions may provide a powerful tool for studying the neural basis of
ToM, especially in combination with computational models of ToM, which offer
quantitative predictions of both when and how much (or how likely) people are thinking
about others’ thoughts. For example, Bhatt, Lohrenz, Camerer, & Montague (2010)
created a competitive buying and selling game in which participants could try to bluff
about the value of an object. The authors predicted that reliance on ToM would increase
in proportion to the riskiness of the bluff, i.e. the discrepancy between the object’s true
value and the proposed price. Consistent with this idea, a region near the right TPJ
showed activity correlated with bluff riskiness across trials. They suggested that
participants may be more likely to engage in ToM, engage in more ToM, or engage in
ToM for longer, when they are making a riskier bluff relative to a less risky bluff, and this
20 of 179
relationship is continuous over a large range of possible risks. Similarly, Behrens et al.
(2008) manipulated the trustworthiness of a confederate over time: the other player
provided “advice” to the participant; this advice shifted over the experiments, so that it
was reliable in some phases, and unreliable in others. The response in ToM regions
tracks with trial-by-trial error in expectations about the informant’s reliability.
Finally, along with manipulating thinking about thoughts across stimuli and tasks, it is
also possible to look at when participants are thinking about thoughts within a single
ongoing stimulus and task. Stories about human actions and beliefs can be broken
down into sections, separating the description of the background and set-up from the
specific sentence that describes or suggests a character’s mental states. Thinking about
other minds can thus be pinned to a specific segment within an ongoing story. The right
temporo-parietal junction, in particular, shows activity at the point within a single story
when a character’s thoughts are mentioned (Mason & Just, 2011; Saxe & Wexler, 2005;
Young & Saxe, 2008). A similar manipulation, dividing a 60-second story into 20-second
segments, only one of which has mental information, has been used in children (Saxe et
al., 2009).
In sum, neuroimaging experiments on understanding other minds produce similar
results, across a wide range of participants, methods, and materials. Similar
experiments have been conducted in Britain, the USA, Japan, Germany, China, the
Netherlands, and Italy (e.g. Leshinskaya & Caramazza, 2014; Moriguchi et al., 2006
Markus van Ackeren, personal communication; Perner, Aichhorn, Kronbichler, Staffen, &
Ladurner, 2006; Schnell et al., 2010). The same brain regions are found in participants
ranging from 5 years old (e.g. Decety et al., 2008; Gweon et al., 2012; Saxe et al.,
2009) to at least 65 years old (e.g. Bedny, Pascual-Leone, & Saxe, 2009; Fletcher et al.,
1995). These regions recruited in adults with high functioning autism (Dufour et al.,
2013), adults who have been completely blind since birth (Bedny et al., 2009); ongoing
work by my colleagues and I suggests they are found in congenitally deaf adults as well.
As described above, the same set of regions responds whether the stimuli is presented
in text or with pictures, visually or aurally. Participants can be thinking about the
thoughts and feelings of a real person or a fictional character.3 The stimuli can include
complex narratives, or just a single thought; participants can be instructed to consider
others’ thoughts and feelings, or be led to do so spontaneously.
The range of tasks, stimuli, and populations make it all the more striking that these
experiments converge on the same conclusions. A consistent group of brain regions
shows increased metabolic activity across all of these experiments in the “mental” or
“theory of mind” condition: regions in bilateral temporo-parietal junction, medial parietal/
precuneus, medial prefrontal cortex, and anterior superior temporal sulcus. Though this
generalization is striking on its own, a key question is: why? Which cognitive process,
invoked by all of these diverse tasks, is specifically necessary and sufficient to elicit
activity in these brain regions?
3
Interestingly, participants can also be assigning hypothetical thoughts and preferences to themselves
(e.g. Lombardo et al., 2010; Vogeley et al., 2001). Note, however, that not all metacognition elicits ToM
activity. The link between attributing hypothetical thoughts to the self, vs. other kinds of metacognition, is
not completely clear (see Saxe & Offen, 2009).
21 of 179
Specificity
During the initial discovery of the ToM brain regions, the first experiments (e.g. Fletcher
et al., 1995; Gallagher et al., 2000) compared two conditions that differed on multiple
dimensions. Compared with the control stories,
the stories about false beliefs also included
more individual characters, more specific
human actions, more implied human emotions,
more invisible causal mechanisms, more social
roles, more unexpected events, more demand
to consider false representations of the world,
different syntax, and so on. In fact, an inherent
risk of such complex stimuli is that there may
be hidden dimensions along which the stimuli
grouped into separate conditions differ and that
it is these differences, rather than the intended
manipulation, that lead to differential brain
activity.
Given all these dimensions, how can we infer
which are the necessary and sufficient features
that led to activity in each region during a given
task? One approach is to try to design an
experiment that contrasts minimal pairs: stimuli
and tasks that differ only in one key dimension,
but are exactly identical on all other
dimensions. Taking this approach, Saxe,
Schulz, and Jiang (2006) were able to match
Figure 1.8: Example stimuli from Saxe et
both the stimulus and the participant’s
al. (2006)
response, using task instructions to change just
how the participants construed the stimulus.
The stimulus was a stick-figure animation of a girl. Modeled after a false-belief transfer
(change of location) task, a bar of chocolate moved from one box to another, while the
girl either faced toward the transfer or away. In the first half of the experiment, rather
than introducing the task as a false-belief task, participants were trained to treat the
stick-figure as a physical cue to the final location of the chocolate bar using three rules,
including the critical Rule 1: ‘Facing = last; Away = first. If the girl is facing the boxes at
the end of the trial, press the button for the last box. If the girl is looking away from the
boxes, press the button for the first box.” (Figure 1.8) Participants were accurate in the
task, but found it difficult and unnatural. In the second half of the experiment,
participants were then told that for Rule 1, another strategy was possible: namely to
view the stick-figure as a person and to consider that person’s thoughts. Rule 1 was
equivalent to judging, based on what the character had seen, where she thought the
chocolate was. Both ways of solving “Rule 1” generate the same behavioral responses,
but only in the second half of the experiment were participants construing Rule 1 as
referring to a character’s thoughts. Saxe et al. (2006) found that, though the stimuli and
the responses were identical across tasks, right TPJ activity was significantly higher in
22 of 179
the second half, when participants were using ToM to solve the puzzle, rather than the
simple association rule.
Another approach to dealing with the many dimensions of ToM stimuli is to
systematically vary or match each of these dimensions in a long series of experiments.
Although each experiment has many differences between the mental and control
conditions, across the whole set of experiments, most other kinds of differences are
eliminated, leaving only one systematic factor: thinking about thoughts.
For example, both Gallagher et al. (2000) and Kobayashi, Glover & Temple (2007)
showed that none of the low-level features of the original verbal stimuli (e.g. number of
nouns, number of straight edges, retinal position) are necessary to elicit activity in these
brain regions, because they found the same patterns of activity in response to verbal
false belief stories and non-verbal false belief cartoons. Within verbal stories, it is not
necessary to explicitly state a character’s thoughts or beliefs: there is activity in these
regions both when people read about a character’s thoughts and when they infer those
thoughts from the character’s actions (Mason & Just, 2011; Young & Saxe, 2009). Nor is
it necessary that the beliefs in question be false: true beliefs, false beliefs, and beliefs
whose veracity is unknown are all sufficient to elicit robust activity in these brain regions
(Döhnel et al., 2012; Jenkins & Mitchell, 2010; Young, Nichols, & Saxe, 2010b).
Other experiments showed that the presence of a human character in the stimuli is not
sufficient: stories that describe a character’s physical appearance, or even their internal
(but not mental) experiences, like hunger or queasiness or physical pain, elicit much
less response than stories about the character’s beliefs, desires, and emotions (Bedny
et al., 2009; Bruneau & Saxe, 2012; Lombardo, Chakrabarti, Bullmore, Baron-Cohen, &
Consortium, 2011; Saxe & Powell, 2006). It is also not sufficient for a story to describe
invisible causal mechanisms (like melting and rusting, Saxe & Kanwisher, 2003),
unexpected events (like a ball of dough that rises to be as big as a house, Young,
Dodell-Feder, et al., 2010b; Gweon et al., 2012), or people’s stable social roles
(including kinship and professional relationships, Saxe & Wexler, 2005; Gweon et al.,
2012), if the story does not also invoke thinking about a person’s thoughts.
One particularly important dimension to test was whether considering any
representation of the world, mental or otherwise, would be sufficient to elicit activity in
these brain regions. Understanding other minds often requires the ability to suspend
one’s own beliefs and knowledge, and consider the world as it would seem from another
perspective. These cognitive processes have been called meta-representation (the
ability to conceive of distinct representations of the world; Aichhorn, Perner, Weiss,
Kronbichler, Staffen, & Ladurner, 2009; Perner, 1991; Perner et al., 2006), and
decoupling (the ability to suspend one’s own knowledge in order to respond from a
different perspective, Leslie & Frith, 1990; Liu, Sabbagh, & Gehring, 2004). Since metarepresentation and decoupling are such essential ingredients of understanding other
minds, and especially understanding false beliefs, many scientists initially hypothesized
that brain regions recruited by false belief tasks most likely performed one of these two
functions. To test this hypothesis, we need stimuli or tasks that require metarepresentation and decoupling, but are not about understanding other minds. Currently,
the best such example are stories about “false signs” and “false maps” (Zaitchik, 1990).
23 of 179
Like beliefs, signs and maps represent (and sometimes misrepresent) reality. Thinking
about the world as depicted in a map requires the capacity for meta-representation;
when the map is wrong, reasoning about the world as it seems in the map requires
decoupling from one’s own knowledge of reality. Nevertheless, stories about false signs
and maps elicit much less activity in these brain regions (especially right TPJ) than
stories about beliefs (Aichhorn et al., 2009; Perner et al., 2006; Saxe & Kanwisher
2003).
In sum, tasks and stimuli that require, or robustly suggest, thinking about thoughts lead
to activity in these regions. Thinking about thoughts can be manipulated by changing
participants’ instructions for the same stimuli, or by changing the stimuli with the same
instructions. Very similar stimuli and tasks, however, which focus on physical objects,
physical representations, or externally observable properties, do not lead to activity in
these regions.
Links to behavior
The review in the previous section shows that tasks and stimuli that evoke thinking
about thoughts also elicit metabolic activity in the ToM brain regions. However, there is
even stronger evidence for a link between activity in these regions and understanding
other minds: across trials, across individuals, and across development, performance on
behavioral tests of ToM is related to brain activity in these same regions.
Most adults pass standard laboratory ToM tasks 100% of the time, leaving little room for
inter-individual variability in accuracy. However, by using tasks with no simple right
answer, it is possible to reveal individual differences in mental state attribution. Imagine,
for example, learning about a girl Grace, who was on a tour of a chemical plant. While
making coffee, Grace found a jar of white powder, labeled “sugar,” next to the coffee
machine. She put the white powder, which was actually a dangerous toxic poison, in
someone else’s coffee, who drank the poison and got sick. Is Grace morally
blameworthy? These scenarios require weighing what Grace intended (her mental
state) against what she did (the outcome), and participants disagree in their judgments;
some people think she is completely innocent (because she believed the powder was
sugar), whereas others assign some moral blame (because she hurt someone). This
difference is correlated with neural activity during moral judgments across individuals;
the more activity there was in a participant’s right TPJ, in particular, the more the
participant forgave Grace for her accidental harms (Young & Saxe 2009b).
An alternative strategy is to measure the quantity and quality of people’s spontaneous
mentalistic attributions to ambiguous stimuli. For example, when viewing the simple
animations of a small and large moving triangle (Castelli et al., 2000), people generate
very rich mentalistic interpretations from the simple movements depicted in these stimuli
(e.g. “the child is pretending to do nothing, to fool the parent”). People differ in the
amount, and appropriateness, of the thoughts and feelings that they infer from the
animations. People who have more activity in ToM brain regions, while watching the
animations give more appropriate descriptions of the triangles’ thoughts and feelings
after the scan (Moriguchi et al., 2006). Relatedly, Wagner, Kelley, & Heatherton (2011)
showed participants still photographs of natural scenes, approximately a quarter of
24 of 179
which contained multiple people in a social interaction. Participants performed an
orthogonal categorization task (“animal, vegetable, mineral?”). Individuals who scored
high on a separate questionnaire, measuring tendencies to think about others’ thoughts
and feelings (the “empathizing quotient”; Baron-Cohen & Wheelwright 2004; Lawrence,
Shaw, Baker, Baron-Cohen, & David, 2004) also showed higher activity in mPFC in
response to photographs of interacting people.
Differences in ToM are easier to find in young children, who are still learning how to
understand other minds. Interestingly, two recent studies from our laboratories suggest
that getting older, and getting better at understanding other minds, is associated,
overall, not with more activity in “ToM brain regions,” but with more selective activity.
Children from 5 to 12 years old all have adult-like neural activity when listening to
stories about characters’ thoughts and feelings. What is different is that ToM regions in
younger children show similarly high activity when listening to any information about
characters in the story, including the characters’ physical appearance or social
relationships (Saxe et al., 2009; Gweon et al., 2012), whereas in older children and
adults, the ToM brain regions are recruited only when listening to information about
thoughts and feelings (Saxe & Powell, 2006; Saxe et al., 2009). This developmental
difference in the selectivity of the ToM brain regions is correlated with age, but also with
performance outside the scanner on difficult ToM tasks (Gweon et al., 2012). Moreover,
the correlation between neural “selectivity” and behavioral task performance remains
significant in the right TPJ, even after accounting for age.
One limitation of all the foregoing studies is that they are necessarily correlational. The
strongest evidence that some brain regions are involved in a cognitive task is to show
that disrupting those regions leads to biases or disruption in task performance.
Transcranial magnetic stimulation (TMS) offers a tool for temporarily disrupting a
targeted brain region. Young, Camprodon, Hauser, Pascual-Leone, & Saxe (2010a)
compared people’s moral judgments following TMS to either right TPJ or a control brain
region 5 cm away. TMS to right TPJ, but not the control region, produced moral
judgments temporarily biased away from considerations of mental state information.
Innocent accidents appeared more blameworthy, while failed attempts appeared less
blameworthy, as though it mattered less what the agent believed she was doing, and
more what she actually did. People didn’t lose the ability to make moral judgments
altogether; they still judged that it is completely morally wrong to intentionally kill, and
not wrong at all to simply serve someone soup. Disrupting the right TPJ thus appears to
leave moral judgment overall intact, but impairs people’s ability to integrate
considerations of the character’s thoughts into their moral judgments. Converging
evidence comes from another TMS study: TMS to right TPJ made adults slower to
recognize a false belief, in a simple (non-moral) false belief task (Costa, Torriero, &
Oliveri, 2008).
Another way to study the necessary contributions of a brain region is to work with
people who have suffered permanent focal (i.e. local) damage to that region, typically
due to a stroke. Samson, Apperly, and colleagues (Apperly & Butterfill 2009; Apperly,
Samson, & Humphreys, 2005; Samson, Apperly, & Humphreys, 2007) have conducted
a series of elegant studies using this approach. Initially, these authors tested a large
group of people, with damage to many different brain regions, on a set of carefully
25 of 179
controlled tasks. They then identified individuals with a specific pattern of performance:
individuals who passed all the control tasks (e.g. measuring memory, cognitive control,
etc.), but still failed to predict a character’s actions based on their false beliefs. Next, the
scientists used a lesion-overlap analysis to ask which brain region was damaged in all,
and only, the patients with this diagnostic profile of performance. The answer was the
left TPJ, one of the same brain regions identified by fMRI4.
Summary
The literature from the last 15 years thus suggests a generalization—there are cortical
regions in the human brain where activity is associated with understanding other minds
in three ways:
1.
Metabolic measures of activity reliably increase when the participant is thinking
about thoughts, across a wide range of stimuli and tasks, but not in response to a
variety of similar control tasks and stimuli.
2.
Activity is correlated with behavioral measures of thinking about thoughts.
3.
Disrupting activity leads to deficits in thinking about thoughts.
So far, these claims are relatively uncontroversial. As noted above, there is a broad
consensus in social cognitive neuroscience. However, much controversy remains about
the proper interpretation of these data.
Exactly what is the nature of these regions, their functions, and their contribution to
thinking about thoughts? Here’s a strong hypothesis: one or more of these regions has
the specific cognitive function of representing people’s mental states and experiences—
that is, of thinking about thoughts. Whenever we are thinking about thoughts, there are
neurons in these regions firing. These neurons are gathered in spatial proximity (i.e. into
a “region”) because they have related computational properties, that are distinct from
the computation properties of neurons in the surrounding cortex.5 The pattern of firing,
in space and time, of these neurons encodes aspects of someone’s thoughts. As an
analogy, consider the way MT neurons encode speed and direction of motion, and face
area neurons encode aspects of facial features that are relevant to face identity
(Freiwald, Tsao, & Livingstone, 2009; Georgopoulos, Schwartz, & Kettner, 1986). In the
proposed hypothesis, the neurons in the ToM brain regions encode aspects and
dimensions of inferred thoughts. Scrambling the pattern of activity in these neurons
would therefore lead to an inability to discriminate one inferred mental state from
another, for example, making all minds appear homogenous: people might all seem to
have the same desires and preferences, and the same knowledge and beliefs. More
4
It is worth noting that the effects of lesions to the right TPJ, one of the regions argued to be most
selective for ToM, haven’t yet been effectively tested. The candidate participants all had extensive and
diffuse damage to the right hemisphere, and failed the control tasks, making it impossible to test ToM
deficits specifically.
5
These distinct properties may derive entirely from patterns of connectivity, not from the structure of the
neurons themselves.
26 of 179
serious damage to these regions might make it impossible to think about other minds at
all, without similarly impairing the rest of cognition.
There certainly is not enough evidence to prove that this strong hypothesis is right; a
more immediate question is whether it is obviously wrong. There are at least two
classes of potential objections: theoretical arguments, based on general principles of
how the brain works, and empirical arguments, based on the results of other
experiments in cognitive neuroscience. In the next section, I describe some of these
objections, and some of our responses to them.
A Strong Hypothesis
Objections from theoretical considerations
Many authors have expressed discomfort with the project of trying to link specific
cognitive functions with delineated brain regions. For example, a decade after their
original chapter, Chris and Uta Frith wrote: “We passionately believe that social
cognitive neuroscience needs to break away from a restrictive phrenology that links
circumscribed brain regions to underspecified social processes” (Frith & Frith, 2012).
Others have echoed this accusation of phrenology; for example, Bob Knight criticizes
the “phrenological notion that a given innate mental faculty is based solely in just one
part of the brain” (Knight, 2007), and William Uttal recently argued that “any studies
using brain images that report single areas of activation exclusively associated with any
particular cognitive process should a priori be considered to be artifacts of the arbitrary
thresholds set by investigators and seriously questioned” (Uttal, 2011). Most such
theoretical objections include variations on three themes: social cognitive
neuroscientists are accused of (incorrectly) viewing regions as (1) functioning in
isolation, (2) internally functionally homogenous, and (3) spatially bounded and distinct.
Here, I address each of these concerns in turn.
First, does claiming that a region has a specific function (e.g. in thinking about thoughts)
entail suggesting that this region functions in isolation? To put it more extremely, does
the hypothesis claim that, for example, the right temporo-parietal junction (RTPJ) could
pass a false belief task on its own? Obviously not. Performing any cognitive task
necessarily depends on many different cognitive and computational processes, and
therefore brain regions. No interesting behavioral task can be accomplished by a single
region. The tasks of the mind and brain—recognizing a friend, understanding a
sentence, deciding what to eat for dinner—must all be accomplished by a long
sequence of processing steps, passing information between many different regions or
computations, from sensory processing all the way to motor action. The function of a
neuron or a brain region should never be identified with completing a cognitive task.
Thus, for example, “passing a false belief task” is not even a candidate function of a
brain region. Any time an individual passes a false belief task, many brain regions—
involved in perceiving the stimuli, manipulating ideas in working memory, making a
decision, and producing a response—will all be required (Bloom & German, 2000).
More generally, functions of regions (or neural population, regardless of spatial
organization) are likely to receive a class of inputs, and transform them into output,
27 of 179
which make different information relatively explicit. Therefore, the specific questions
about any neural population should include: what input does it receive, what output
does it produce, and what information is made explicit in that transformation? Of course,
the answers to these questions will require us to understand the position of this neural
population within a larger network, especially when characterizing a region’s input and
output. In the case of the ToM regions, a related question concerns the relationships
between the different regions within the network. At least five cortical regions are
commonly recruited during many different social cognitive tasks: how is information
passed between, and transformed by, each of these spatially distinct regions?
Thus, studying the function of a brain region means studying in isolation one component
of a system that could never function in isolation. This description may sound ominous,
but scientific progress frequently requires us to break complex systems into component
parts. While the pieces could not function in isolation, understanding their isolated
contributions is necessary to understanding the function of the integrated system. For
any given neural population, it is reasonably to ask: what classes of stimuli and tasks
predictably and systematically elicit increased activity in the population as a whole?
Which dimensions of stimuli lead to activity in distinct subpopulations of neurons? Both
traditional and new fMRI methods help answer these questions, albeit somewhat
indirectly (more on this, in What Now?).
The second objection is that studying brain regions leads to the false assumption that
groups of spatially adjacent neurons are functionally homogenous. The regions we
study in fMRI are orders of magnitude larger than the true computational units of brain
processing, the neurons. Changes in blood oxygenation measured by fMRI inevitably
reflect averages over the activity of many thousands of individual neurons. Why is it
useful to study oxygen flow to chunks of cortex approximately 1–5 cm2 in size, which
are so much bigger than neurons, but so much smaller than the networks required to
complete a task?
Our suggestion is that there is no reason, a priori. It just happens, as a matter of
empirical fact, that many interesting computational properties of the brain can be
detected by studying the organization of neural responses on this scale. Aggregating
the responses of neighboring neurons often produces informative population averages.
Results obtained via fMRI reflect the same distinctions found in directly observable
population codes (e.g. Kamitani & Tong, 2005; Kriegeskorte, Bandettini, & Bandettini,
2007). This empirical fact may have a theoretical explanation. Neurons with similar or
related functions may be spatially clustered to increase the computational efficiency of
frequent comparisons (e.g. lateral inhibition). Blood-oxygen delivery to the cortex may
follow the contours of neural computations, to increase the hemodynamic efficiency of
simultaneously getting oxygen to all of the neurons that need it (Kanwisher, 2010).
However, these arguments are not necessary premises; for fMRI to be useful, we only
need the empirically observable fact that useful and reliable generalizations can be
made for hemodynamic activity in patches of cortex at the spatial scale of millimeters.
Of course, though, neurons within a region or an fMRI voxel are never functionally
homogenous. Consider the analogy of primary visual cortex—neurons in V1 have visual
receptive fields, meaning that activity can be induced by a pattern of bars of light falling
28 of 179
on a specific region of the retina. However, neurons in V1 differ from one another in
where on the retina one should place the light (retinotopy), how large the pattern should
be (size and spatial frequency preferences), and the orientation to which the bars
should be rotated (orientation selectivity), to elicit a maximal response. There is also a
separate (but systematically interleaved) population of neurons for which the response
depends on color, but not orientation or size. Furthermore, some neurons primarily send
information to subsequent regions of visual cortex (e.g. excitatory neurons) whereas
other neurons primarily modulate the response of neighboring neurons in V1 (e.g.
inhibitory interneurons). As far as we know, the metabolic activity measured by fMRI
reflects a combination of activity in all of these different populations. Consequently, we
should never assume that the amount of “activity” we measure in a region with fMRI
represents (or would correlate very well with) the rate of firing of any individual neuron
inside that region. Similarly, we cannot assume that if two different stimuli or tasks elicit
similar magnitudes of activity in a region, then they are eliciting responses in the same,
or even shared, neural populations. Completely non-overlapping subpopulations of
neurons could produce the same magnitude of fMRI activity within a region. Any
interpretation of fMRI data must be sensitive to this possibility. In fact, studying the
organization of functional subpopulations within a region (e.g. which dimensions of
stimuli are represented by distinct subpopulations within a region) may be one of the
most powerful ways that fMRI will contribute to the neuroscience of ToM. I describe
these methods in greater detail in What Now?.
The third potential objection is that studying regions with fMRI leads researchers to
imagine boundaries between discrete regions, when the truth is a continuous
distribution of neural responses over the cortical sheet. The data we described in the
first section shows that cortex is not functionally homogenous with regard to theory of
mind, and regional distinctions are not all “artifacts of arbitrary thresholds.” Still, there is
a legitimate reason why cognitive neuroscientists may be reluctant to call any reliable
functional regularity discovered by fMRI a “region.” These “regions” may turn out to be
just one piece of a larger continuous functional map over cortex, not computationally
distinct areas of their own (Kriegeskorte, Goebel, & Bandettini, 2006).
Cortical responses at scales measurable by fMRI are organized along multiple
orthogonal spatial principles. One is the division of cortex into cytoarchitectonic “areas,”
like primary visual cortex (V1) and primary auditory cortex (A1). Orthogonal to the
division of cortex into areas are topographic principles. Most visual regions, for
example, are organized by retinotopy: moving across the cortical sheet, the region of
the visual field eliciting a maximal response varies smoothly, covering the whole visual
field from fovea to periphery, top to bottom, and left to right. Likewise, multiple distinct
motor areas are organized by somatotopy, and auditory areas by tonotopy.
These orthogonal principles of cortical organization create a challenge for cognitive
neuroscientists, because in charting new territory, away from well-understood sensory
and motor systems, we may claim to discover new functional regions associated with
higher-order cognitive processes, which are really just one end of a larger map (cf.
Konkle & Oliva, 2012). To make the puzzle concrete, imagine looking at functional
responses to visual stimuli across occipital cortex for the first time, without the benefit of
the history of visual neuroscience. One tempting way to divide the occipital cortex into
29 of 179
functional “regions” might be by retinotopy—one group of patches that responds to
foveal stimuli, and a different group of patches that responds to peripheral stimuli. This
foveal vs. peripheral difference is highly robust, replicable within and across subjects,
within and across tasks, and correlated with behavior (i.e. visual performance in
corresponding regions of the visual field). Nevertheless, other considerations, such as
cytoarchitecture, connectivity, and processing time, suggest that this is the wrong
division for capturing functional and computational regularities. Identifying a robust
functional regularity that divides one patch of cortex from another is not the same thing
as identifying a true cortical area—a region that is computationally distinct from its
neighbors, with distinct cytoarchitecture, connectivity, and topography (Friston, Holmes,
Worsley, Poline, Frith, & Frackowiak, 1995; Kanwisher, 2010; Kriegeskorte et al., 2006;
Worsley, Evans, Marrett, & Neelin, 1992). Instead of studying the “peripheral patches,”
neuroscientists divide the cortex into V1, V2, V3, MT, etc., each of which is then
internally organized (in part) by retinotopy.
Imagine you are a social cognitive neuroscientist, looking at a new bit of cortex, and you
see a new functional regularity—a patch of cortex that shows a high response when the
individual is thinking about stories, cartoons, or movies depicting the contents of another
mind. Which kind of functional regularity is this? On the one hand, these data could
signal the discovery of a true computational area, like V1. On the other hand, the
observed functional regularity might be more like “peripheral patches,” one part of a
stimulus space or dimension that is mapped across cortex, but crosscuts multiple
computational areas.
We believe that current fMRI data cannot resolve this puzzle directly. One approach
may therefore be to suspend judgment until other sources of evidence are available. If
patches of cortex involved in understanding other minds are true cortical areas, it will be
possible to distinguish them from their anatomical neighbors by cytoarchitecture,
connectivity, and/or topography. Non-invasive functional imaging tools do not yet have
high enough resolution to reveal cytoarchitecture in vivo. To study the links between
function and cytoarchitecture with current technology, we would need to collect
functional data to identify the regions in vivo, and then analyze the neuroanatomy of the
same individual post mortem.
A weaker, but more accessible, source of evidence is patterns of connectivity. In some
cases, cortical areas can be differentiated by their profiles of connectivity. It has become
increasingly possible to measure the pattern of connections between regions using
neuroimaging. Diffusion imaging (DTI), which looks at the predominant direction of
water diffusion, allows us to visualize the dominant pathways of axons connecting brain
regions. Using diffusion, we can ask whether a patch of cortex involved in
understanding other minds shows a different pattern of connectivity than its neighbors to
the rest of the brain. Initial evidence suggests it does, at least in the case of the right
TPJ. Mars, Sallet, Schüffelgen, Jbabdi, Toni, & Rushworth (2012) found that the broad
area of right temporo-parietal cortex (BA 39/40) can be sub-divided into three clusters,
based on DTI connectivity alone, and one of these clusters (the posterior one) is
functionally correlated (during a resting baseline) with the other ToM regions, including
medial prefrontal and medial parietal cortex. Bzdok et al. (2013) corroborated and
extended this finding, using resting state connectivity and meta-analyses. They found
30 of 179
similar anatomical parcellation, and also tested functional profile: the region of TPJ that
correlated with other ToM regions was also the region that responded most to social
cognition tasks. Similarly, in a recent meta-analysis, Schurz et al. (2014) found that the
same posterior region of TPJ defined by tractography responded more to false belief
tasks that any other sort of social cognition task (such as trait inference, violations of
rational action, or strategic games); while the region anterior to it (posterior STS)
responded most to violations of rational actions. In contrast, regions MPFC respond
strongly to trait inference tasks.
Currently, however, neither cytoarchitecture nor connectivity analyses give a definitive
evidence that patches of cortex recruited by mental state reasoning tasks are true
cortical areas. An alternative approach might therefore be to consider the alternative
hypothesis directly—these regions are one part of a larger, continuous map. If we
compare the cortical patches we are studying to their anatomical neighbors, is there a
plausible higher-level “stimulus space,” which could unite these responses into one
map?
Again, the right TPJ is an interesting example. The right TPJ region that is activated by
ToM tasks (as described in “Theory of mind and the brain”) has two very close
anatomical neighbors. One neighbor (up toward parietal cortex in the right inferior
intraparietal sulcus (IPS), but confusingly sometimes also called RTPJ) is recruited by
unexpected events that demand attention. These events may be unexpected because
they are rare, or because a generally reliable cue was misleading on this occasion.
Redirecting attention toward an unexpected event leads to metabolic activity in this
region (Corbetta & Shulman, 2002; Mitchell, 2008; Serences, Shomstein, Leber, Golay,
Egeth, & Yantis, 2005). Damage to this region makes it hard for objects and events in
the contralateral visual field to attract attention, producing left hemifield neglect
(Corbetta, Patel, & Shulman, 2008). Some authors have noted that false belief tasks
typically involves an unexpected event (Corbetta et al., 2008; Decety & Lamm, 2007;
Mitchell, 2008): for example, false belief tasks often hinge on an object unexpectedly
changing location while the protagonist is out of room, and require the participant to shift
attention between both locations. In fact, the region that is recruited during exogenous
attention tasks is so close to the region recruited during ToM tasks that some concluded
that these are actually just two different ways to identify the same region (Corbetta,
Kincade, Ollinger, McAvoy & Shulman, 2000; Mitchell, 2008). More recently, however,
connectivity analyses (Bzdok et al., 2012; Krall et al., 2014; Mars et al., 2012), metaanalyses (Schurz et al, 2014; Decety & Lamm, 2007; Kubit & Jack, 2013), highresolution scanning within individual subjects (Scholz, Triantafyllou, Whitfield-Gabrieli,
Brown, & Saxe, 2009) suggest that there are actually two distinct cortical regions (or
“patches”), and the region recruited by attention tasks is approximately 10 mm superior
to the region recruited by ToM tasks (Krall et al., 2014; Scholz et al., 2009). Also, the
region that is recruited during false belief tasks is not recruited by unexpected transfers
of location in stories about false maps and physical representations, as described above
young (Young, Dodell-Feder, & Saxe, 2010a). Nevertheless, it remains an interesting
possibility that the exogenous attention response and the ToM response are part of a
larger continuous map across cortex, a topography of different kinds of unexpected
attentional shifts. The more superior end of this map could direct attention toward
31 of 179
unexpected positions in space and time, while the more inferior end of the map could
direct attention toward unexpected people, actions, or inferred thoughts.
A second topographical “stimulus space” of which the RTPJ could be a part runs
anteriorly, through the right superior temporal sulcus (STS), and down toward the
temporal pole. As with the RTPJ and the right IPS, the RTPJ and the posterior STS
were initially conflated, but have subsequently been shown to be functionally distinct
(Perner & Leekam, 2008; Perner & Roessler, 2012; Schurz, Radua, Aichhorn, Richlan,
& Perner, 2014; Gobbini, Koralek, Bryan, KJ, & Haxby, 2007). Multiple parts of the STS
are recruited when participants observe other people’s actions in photographs, film
clips, point light walkers or animations (Pelphry, Mitchell, McKeown, Goldstein, Allison,
& McCarthy, 2003; Pelphrey, Morris, Michelich, Allison, & McCarthy, 2005; Hein &
Knight, 2008). The STS is large, and Kevin Pelphrey and colleagues (Pelphrey et al.,
2005; Pelphrey & Morris 2006) propose that it contains a pseudo-somatotopic map of
observed actions: others’ mouth motions represented most anteriorly, followed by hand
and body movements, followed by head and eye movements represented most
posteriorly. An intriguing possibility is that the temporo-parietal junction, which is at the
most posterior end of the superior temporal sulcus, is part of this same map. Whereas
hand and body movements convey information about what a person is doing and
intending, head and eye movements convey information about what a person is looking
at and seeing. Thus, the STS may contain a map of others actions that move from
externally observable body movements (anterior end) toward invisible mental states
(posterior end), culminating in the RTPJ, which responds to thinking about what a
person is thinking.
In principle, either of these mapping hypotheses, or both, could be true. Evidence that a
ToM region, e.g. the RTPJ, is part of a larger cortical map might come in the form of a
response that moves continuously across contiguous patches of cortex, modulated by
continuous changes in a stimulus dimension (Konkle & Oliva, 2012). For example, if the
RTPJ is one end of an attention map, we might expect to see that surprising facts about
physical entities elicit activity relatively far from the TPJ, but that as the unexpected
stimulus becomes more social and interpersonal, activity moves continuously"
across the map, ending at the RTPJ. Similarly, if the RTPJ is the “abstract” or “head”
end of a map of observable social actions, we’d expect to see continuous changes in
the location of activity, as depictions of human actions become either more abstract or
physically higher on the body. Finally, both could be true. The RTPJ could exist at its
precise location because that is where a region that deals with unexpected information
converges with a region that deals with human actions. We would be enthusiastic to see
someone test these hypotheses further.
For now, however, we suggest that it does not matter for any of our empirical purposes
whether the RTPJ (or any other patch of cortex recruited in ToM tasks) is a true cortical
area, or whether it is one end of a larger map, as defined by cytoarchitecture,
connectivity, and topography. In either case, there is a robust functional regularity
replicable within and across subjects, within and across tasks, and correlated with
behavior: a bounded, adjacent patch of cortex where activity is high during a range of
ToM tasks. Without committing to whether they are true computational “areas,” or just
the equivalent of “Peripheral Patches,” we believe that it is fruitful to continue to study
32 of 179
these coherent patches as “regions,” in order to discover the computations and
representations that underlie thinking about other minds.
Given everything else we know about the brain, it is not surprising that systematic
response profiles are linked to regions of cortex much bigger than a neuron and much
smaller than a network. The challenge now is to integrate the huge, and growing, list of
empirical discoveries and to construct hypotheses about the computations and
representations in the neural populations we are studying. These empirical results form
the basis of a different set of potential objections to the hypothesis that there is a strong
specific link between cortical regions, like the TPJ, mPFC, and PC, and thinking about
thoughts.
Empirical objections
An empirical objection to our argument in “Theory of mind and the brain” might begin by
pointing out that our review of the literature was selective. In addition to the dozens of
articles we cited, there are dozens of others that claim so-called “ToM regions” are
active in tasks that do not involve thinking about thoughts, or are not active in tasks that
do involve thinking about thoughts. How can we integrate these other data into a
coherent hypothesis?
First, what about claims that “ToM regions” are active in tasks that do not involve
thinking about thoughts? The RTPJ is again a useful example. As we mentioned above,
some authors initially believed that the same region of RTPJ involved in false belief
tasks was also recruited during any exogenous shift of attention and/or biological motion
perception. Other literature suggests that the RTPJ is involved in maintaining a
representation of one’s own body, by integrating multi-sensory information and locating
the body in space (Blanke et al., 2005; Blanke & Arzy, 2005; Tsakiris et al., 2008).
Experiments that ask people to mentally rotate their own body or imagine their body in
different parts of space, as well as those that induce changes in bodily self-perception
(with e.g. rubber hands) find activation in this region (Arzy et al., 2006; Blanke & Arzy,
2005). TMS and intracranial stimulation to this region has been shown to lead to out-ofbody experiences, confusion of the body vs. the environment, and illusory changes in
the orientation of body parts (Blanke et al., 2002, 2005; Tsakiris et al., 2008).
How do we interpret these results, which seem to contradict our theory? In general, we
could consider five possibilities. The first is that one group of scientists has made an
error, leaving an unnoticed confound in their experimental paradigm. Could the body
representation tasks also involve thinking about thoughts? Could the false belief task
accidentally induce updating maintaining a representation of one’s own body? These
are empirical questions, but we consider both possibilities highly unlikely. The second
option is that there is some deep common computation, served by the same neural
population that is required by these different classes of tasks. One option here would be
“decoupling” (Gallagher & Frith, 2003; Leslie & Frith 1990; Liu, Sabbagh, & & Gehring,
2004), maintaining distinct representational reference frames, e.g. for one’s own and
other minds, for current vs. imaginary body positions, and so on. The third option is that
the same neurons can assume distinct and even unrelated functional roles, depending
on the context and the pattern of activity in other neural populations (e.g. Miller &
33 of 179
Cohen, 2001). On this view, thinking about thoughts and maintaining a body
representation are cognitively unrelated, in spite of being implemented by the same
neurons. The fourth option is that distinct neuronal populations are involved in thinking
about thoughts and maintaining a body representation, but these neurons are
interleaved within the RTPJ, just as color- and orientation-sensitive neural populations
are interleaved in V1.
Finally, the fifth option is that distinct neuronal populations are involved in thinking about
thoughts vs. maintaining one’s body representation, exogenous attention shifts, and
biological motion perception. These neural populations are not interleaved: they are
contained in distinct regions that are merely nearby on the cortical sheet. We find this
last option most plausible. Standard fMRI methods, which involve extensive spatial
blurring at three stages (acquisition, preprocessing, and group averaging; Fedorenko &
Kanwisher, 2009; Logothetis, 2008), are strongly biased to conflate neighboring regions
that are truly distinct. Even so, the existing evidence suggests that the region involved in
body representations is lateral (MNI x-coordinates typically around 64mm) to the region
involved in thinking about thoughts (x-coordinates around 52 mm). As described above,
this was also true of the regions involved in biological motion perception (Gobbini et al.,
2007) and exogenous attention (Decety & Lamm, 2007; Scholz et al., 2009); these
almost completely non-overlapping in individual subjects.
A different kind of challenge arises from examples of tasks that apparently do involve
thinking about thoughts, but do not elicit activity in TPJ, mPFC, or PC. Two interesting
examples are visual perspective-taking tasks, and recognition of facial expressions of
emotions from photographs. In visual perspective taking tasks, the participant sees an
image of a character in a 3D space, and is asked to imagine the view of the room from
the viewpoint of the character. For example, participants may be asked to report how
many dots the character can see (the third person perspective), versus how many dots
the participants themselves can see (including those that are out of the character’s view,
the first person perspective). This perspective-taking task clearly involves thinking about
the character’s visual access, which could be construed as a mental state.
Nevertheless, this task typically does not elicit activity in the same regions as thinking
about thoughts (Aichhorn et al., 2006; Vogeley May, Ritzl, Falkai, Zilles, & Fink, 2004).
Similarly, participants in hundreds of fMRI experiments have viewed photographs of
human faces expressing various basic emotional expressions (e.g. sad, afraid, angry,
surprised, happy, neutral). Although these images do depict evidence of another
person’s emotional experience, they also typically do not elicit activity in the same
regions as thinking about thoughts (Costafreda et al., 2008; Lamm, Batson, & Decety,
2007; Vuilleumier et al., 2001).
What should we conclude from these examples? One option is to make a forward
inference. Using the evidence from these tasks to change our hypothesis about the
brain regions’ functions. Here, the forward inference might be that the ToM brain regions
are responsible for only a subset of mental state processing. We could conclude that
different brain regions are involved in thinking about different classes of internal
experiences: bodily states, emotional states, perceptual states, or epistemic states (e.g.
thinking, knowing, doubting, etc.). The so-called ToM brain regions might be specifically
34 of 179
involved in representing epistemic states, while regions of insula represent others’
emotions and regions of parietal cortex represent others’ perceptual states.
Another option is to make a reverse inference: using the pattern of neural activity to
change our analysis of the cognitive processes required by the task. Reverse inferences
are risky because they require a lot of confidence in the functional specificity of the brain
region(s) involved (Poldrack, 2006a). However, we think ToM brain regions are good
candidates to support reverse inference, given the converging evidence across the
many experiments described above. In this case, a reverse inference might be that
these visual perspective taking and emotional face tasks do not actually elicit thinking
about thoughts, and instead are solved by alternative computational strategies. For
example, rather than truly considering someone’s perceptual experiences, line-of-sight
tasks may be solved using mental rotation and geometric calculation. Emotional facial
expressions may be recognized (akin to object recognition) without always requiring a
representation of the person’s internal state.
In these particular cases, we are open to either the forward or reverse inference; only
further experimentation will tell which is the better generalization. In general, we believe
that in this domain, inferences can be made in both directions, forward and reverse. The
absence of activation in perspective taking and emotion recognition tasks provides an
important constraint on the possible functions of ToM brain regions (the forward
inference). At the same time, the fact that visual perspective-taking and emotion
recognition rely on different brain regions from reading stories about thoughts provides
evidence that these tasks depend on different cognitive processes (the reverse
inference). It’s the give and take of these two kinds of inferences, as evidence
accumulates, that allows us to build a coherent understanding of both cognitive and
neural function.
Summary
Taken together, these first two sections illustrate the main contributions of the first
decade of neuroimaging the understanding other minds. We have discovered a robust,
replicable functional regularity in the human brain: regions that have increased activity
when participants think about thoughts. These regions may be true cortical areas or
parts of larger topographical maps, but in either case, understanding other minds is a
major organizing principle of responses over cortex. The function of these regions is not
to complete a task, but to transform some class of input into some output; and the class
of input has something to do with thinking about minds, and not bodies or abstract
representations. Of course, this description remains unsatisfying and largely
underspecified. Nevertheless, we are optimistic that the second decade of this research
program will continue to improve our specifications. In particular, we are excited about
newly emerging methods for fMRI data analysis that focus on the second aspect of a
region’s function: the features within its preferred stimulus class that organize differential
responses within each of the ToM regions.
35 of 179
What Now?
Differences between theory of mind regions
One step in specifying the computations performed by ToM regions will be
understanding the division of labor and information transfer between the different
regions. Overall, ToM regions show similar profiles to most of the contrasts we
described. However, research in the last 5 years has begun to tease apart the functional
profiles of these regions, and the differences are intriguing, though much work still
remains to be done to form a coherent view of how they all fit together.
The most striking contrast comes from a task that elicits very robust activity in the
medial ToM regions (mPFC and precuneus), but not the lateral ToM regions (TPJ and
anterior STS): thinking about personality traits, especially of the self and close others
(Whitfield-Gabrieli et al., 2011; Moran et al., 2011a; Saxe, Moran, Scholz, & Gabrieli,
2006a; Krienen, Tu, & Buckner, 2010). In a typical version of the experiment,
participants in the scanner see single words describing personality traits (e.g. “lazy,”
“talkative”, “ambitious”) and judge either whether each one is desirable or undesirable
(the semantic control condition), is true of a famous person (the other control condition),
or is true of themselves (the self condition). MPFC and precuneus regions show much
higher activity during the self condition; moreover, within the self condition, activity in
mPFC is higher for the words that participants say are true of themselves (Moran et al.,
2011a), and the amount of activity in mPFC for each item presented in the self task (but
not in the semantic or other tasks) predicts participants’ subsequent memory for those
items on a surprise memory test (Mitchell, Macrae, & Banaji, 2006; Jenkins & Mitchell,
2009). These data are compelling to us: the mPFC, but not the TPJ, is involved in
reflection about one’s own stable traits and attributes (Lombardo et al., 2010). Similarly,
elaborating one’s own autobiographical memories leads to activity in medial ToM
regions, whereas imagining someone else’s experiences on similar occasions elicits
activity in bilateral TPJ (Rabin, Gilboa, Stuss, Mar, & Rosenbaum, 2010).
In fact, coding information in terms of similarity to the self may be a key computation of
mPFC. In one series of studies (Tamir & Mitchell, 2010), participants judged the likely
preferences of strangers (e.g. is this person likely to “fear speaking in public” or “enjoy
winter sports”) about whom they had almost no background information. Under those
circumstances, the response of the mPFC (but not TPJ) was predicted by the
discrepancy between the attributions made to the target and the participant’s own
preference for the same items: the more another person was perceived as different from
the self, for a specific item, the larger the response in mPFC.
Another distinction, supported by multiple studies, suggests that sub-regions of mPFC
are most recruited when thinking about someone’s negative emotions or bad intentions,
whereas the TPJ makes no distinction based on valence. For example, Bruneau, Pluta,
& Saxe (2011) found that only the mPFC showed a higher response to stories about
very sad events (e.g. a person proposes marriage and is rejected) compared to neutral
or positive events (e.g. the marriage proposal is accepted). In a PET study, Hayashi,
Abe, Ueno, Shigemune, Mori, Tashiro, et al. (2010) found that a region in mPFC was
recruited when people were considering an actor’s dishonesty as a factor in moral
36 of 179
judgments; and in an fMRI study, Young & Saxe (2009a) found that a region in ventral
mPFC was correlated with moral judgments of attempted harms, which are morally
wrong only because of the actor’s negative intentions. Converging with these
neuroimaging studies, lesion studies suggest that focal damage to ventral mPFC
creates disproportionate difficulty in understanding bad intentions, and in integrating
those intentions into moral judgments (Koenigs et al., 2007).
In this vein, further work has been done to examine the response profile of mPFC. The
“regions” implicated in ToM are very large, especially in the mPFC, and there may be
multiple sub-divisions, each with different response profiles. Proposed distinctions along
the ventral-dorsal axis of mPFC include: similarity to self (such that self-relevant
processes elicit responses more ventrally (e.g. Mitchell et al., 2006; Jacques. Conway,
Lowder, & Cabeza, 2011), interpersonal closeness (people who are closer, or more
important to the self elicit responses more ventrally, Krienen et al., 2010), or affective
content (“hot” affective states elicit responses more ventrally (Ames, Jenkins, Banaji, &
Mitchell, 2008; D’Argembeau et al., 2007; Mitchell & Banaji, 2005), while “cool” cognitive
states elicit responses more dorsally (Kalbe, Schlegel, Sack, Nowak, Dafotakis,
Bangard, et al., 2010; Shamay-Tsoory & Aharon-Peretz, 2007; Shamay-Tsoory, Tomer,
Berger, Goldsher, & Aharon-Peretz, 2005). Thus, while there may be a convergent
theoretical account of the mPFC, and its responses to other people’s negative intentions
and emotions, one’s own personality traits, and ambiguous inferences about
preferences, another possibility is that these response profiles reflect distinct subregions within mPFC, each contributing a distinct computation to understanding other
minds.
Intriguingly, none of these distinctions have shown to affect the lateral ToM regions. In
contrast, one dimension that seems to influence the magnitude of response in TPJ more
than mPFC is whether someone’s thought or feeling is unexpected, given the other
information you have about that person. Saxe & Wexler (2005) introduced characters
whose social background was mundane (e.g. New Jersey) or unusual (e.g. a
polyamorous cult). Participants then read about that character’s thoughts and feelings
(e.g. a husband who believed it would be either fun or awful if his wife had an affair). On
their own, neither the background nor the content of the belief affected the magnitude of
response in the TPJ, but there was a significant interaction: whichever thought was
unlikely, given the character’s social background, elicited a larger response in the right
TPJ. Recently, Cloutier, Gabrieli, O'Young, & Ambady (2011) provided a conceptual
replication of this result: participants saw photographs of people labeled as Democratic
or Republican, paired with opinions that were either typical of their political affiliation or
typical of the opposite affiliation. Opinions that were unexpected given the protagonist’s
political background (e.g. a Republican wanting liberal Supreme Court judges) elicited a
higher response in most of the ToM regions, including bilateral TPJ and mPFC. Finally,
a third study suggests that the conflict between background and belief is necessary for
increased activation, not just sufficient. In the absence of specific background
information about the believer, there is no difference in the response of any ToM region
to absurd vs. commonsense beliefs (e.g. “John believes that swimming in the pool is a
good way to grow fins/cool off,” Young, et al., 2010b).
37 of 179
Although they are preliminary, we find these results exciting because they are
consistent with the idea that activity in TPJ reflects a process of forming a coherent
model of another’s mind. We expect other people to be coherent, unified entities, and
strive to resolve inconsistencies with that expectation (see Hamilton & Sherman, 1996).
These results suggest that while TPJ activity may be related to the discrepancy
between a thought or feeling and other information about the protagonist, mPFC and
PC activity may be related to discrepancies between the protagonist’s thoughts or
feelings and the participant’s own thoughts or feelings on the same topic (see Chapter 5
for more discussion of this idea). Thus, whereas TPJ may be involved in integrating a
belief or preference into a coherent model of another’s mind (e.g. Young et al., 2010c),
mPFC and precuneus activity may reflect a different “anchor-and-adjust” strategy, that
helps identify other people’s thoughts and preferences by starting with one’s own
preferences, and then adjusting them as necessary (Tamir & Mitchell, 2010). Taken
together, these results suggest that medial and lateral ToM regions support distinct
computations within ToM. These distinctions help us separate ToM into its real (i.e.
neurally-realized) component parts, and formulate hypotheses about each of the more
specific functions of individual regions within the group.
Information within theory of mind regions
An equally important step in specifying the computations performed by ToM regions will
be understanding what is happening inside each region. Until now, we asked simply
whether ToM brain regions do or do not show activity in response to a task or stimulus.
However, a single “region” millions of neurons: activation within a region is clearly
neither binary nor uniform.
Patterns within theory of mind regions
One strategy is to look for divisions of functional responses across the neural
populations within each region. Two relatively novel methods for analyzing fMRI data
may allow neuroscientists to look inside ToM regions. Distinctions between neural
subpopulations within regions may provide clues to how these regions function.
The first method is repetition-suppression, also called functional adaptation. Repetitionsuppression analyses take advantage of the observation that after processing a
stimulus or task once, activity in a neuron or brain region in response to an identical
stimulus or task is suppressed, or adapted (Grill-Spector, Henson, & Martin, 2006). By
manipulating the features of the repeated stimulus, so that some are identical and some
are different from the original stimulus, it is possible to ask what counts as the same
stimulus for a particular brain region (Kourtzi & Kanwisher, 2001). If the repeated
stimulus is effectively the same with regard to the features represented by the brain
region, then the region’s response will be suppressed or adapted. On the other hand, if
a feature that is represented in the brain region has been sufficiently modified, the brain
region’s response will “recover” from adaptation.
This method has been used frequently and effectively to study visual representations of
objects and places (Poldrack, 2006b). One a few studieshave taken advantage of this
approach to test hypotheses about ToM. Jenkins, Macrae, & Mitchell (2008) asked
38 of 179
whether attributing a preference (e.g. “enjoys winter sports”) to oneself, a similar
stranger, or a dissimilar stranger, depend on the same neural subpopulations within
mPFC. They found that thinking about a similar other person after thinking about the
self led to repetition suppression, while thinking about the self and then a dissimilar
other led to recovery from adaption. These results support the hypothesis that the
mPFC represents other minds in terms of their similarity to the self. Similarly, Ma and
colleagues (Ma et al., 2013a; Ma, Baetens, Vandekerckhove, Van der Cruyssen, & Van
Overwalle, 2013b) find evidence of representation suppression in MPFC to stimuli that
suggest similarly personality traits, but not repeated object traits.
Currently, we are using a similar strategy to investigate the components of mental state
attributions. People read short stories in which a key mental state was repeated twice,
with some elements changed. After the first mental state sentence (e.g. “Megan thinks
that Julie is being too flirty”), the repetition either changed the agent (e.g. "Gina thinks”),
the attitude verb (e.g. "Megan worries that"), the content (e.g. “Megan thinks that Julie
should be more flirty"), all three, or none of these elements. In preliminary data, we find
that ToM brain regions recover from adaptation for each kind of change on its own
(compared to no change), suggesting that these regions encode all three elements of a
mental state attribution. Interestingly, while the mPFC shows the most recovery when
the content of the mental state changes, the left temporoparietal junction (LTPJ) shows
the most recovery when the attitude verb changes, and the RTPJ shows equal recovery
for any kind of change. If these results hold up to further analyses, they may provide
clues about the contributions of each ToM brain region to thinking about the minds of
others.
The second method, multi-voxel pattern analysis (MVPA), looks for subtle, but reliable
spatial patterns within a single region (or local neighborhood) of cortex. By looking at
spatial separability, MVPA provides a more direct measure of the existence of
functionally separable sub-populations of neurons than repetition suppression. Each
“voxel” (the fMRI equivalent of a pixel) may have some, possibly very small, preference
for one kind of stimulus over another due to biases in the neuronal populations toward
one type of information-processing vs. another; pattern analyses measure the
similarities and differences between these patterns across space (Haxby, 2001). A few
studies have begun to use pattern analyses to study ToM (e.g. Gilbert et al., 2008). In
one promising example, Peelen, Wiggett, & Downing (2006) found that the pattern of
activation in the mPFC reliably reflected the content of another person’s emotion (e.g.
sad vs. angry), independent of the stimulus modality (e.g. vocal expressions vs. body
posture). In Chapters 2, 3, and 4, I show that this technique can used to find
distinctions which represented within a region, and point toward the underlying
distinctions and computations in ToM.
Magnitude: “more” theory of mind
As well as asking how information is distributed across a region, is it equally important
to understand what drives activity, in a region, in a voxel, and in a neuron. The
magnitude of activity, over tasks or stimuli, could reveal not only what class of stimuli is
processed in a region, but also which dimensions of those stimuli and tasks elicit more
or less processing.
39 of 179
Interestingly, initial attempts to modulate activation in the ToM network mostly
discovered features that do not elicit differential magnitudes of response. For example,
most of the ToM regions show an equally high response to explicit descriptions of
beliefs that are true or false (Young & Saxe, 2008), justified or unjustified (Young,
Nichols, & Saxe, 2010c), and well-intentioned or bad (i.e. a girl who believes she is
putting poison in her friends coffee, vs. believes that she is putting sugar in the coffee,
Young, Cushman, Hauser, & Saxe, 2007). In the TPJ, at least, it also does not matter to
whom the thought or feeling is attributed: there is an equally strong response to beliefs
attributed to similar or dissimilar others (Saxe & Wexler, 2005), or to members of one’s
own group vs. an enemy group (Bruneau & Saxe, 2010; Bruneau, Dufour & Saxe,
2012).
Part of the reason for this lack of success may be that we do not yet have satisfactory
cognitive or computational theories of ToM that allow us to predict when “more” ToM
processing will be required, or even exactly what “more” means. Some intuitive
possibilities have already proven empirically false. For example, making it harder to
infer what a character believes, by making the available evidence more ambiguous,
does not lead to more activity in TPJ regions (Jenkins & Mitchell 2009). In fact, we
(Dodell-Feder et al., 2011) found that, while some stories about thoughts systematically
elicit more activity than others, in each ToM region, we could not find any feature (e.g.
vividness, length, syntactic complexity) that predicted these differences in activity, with
the partial exception of the precuneus, which showed greater activity to stories that
involved more people.
Researchers with a background in computer science or game theory often suggest one
particular dimension for “more” ToM processing—the depth of embedding of one mental
state within another. Thus, many people intuit, reasoning about an embedded belief
(e.g. “Carla believes that Ben thinks that she eats too much junk food”) should require
more ToM processing that reasoning about a simple belief (e.g. “Carla believes that she
eats too much junk food.”) When we tested this hypothesis directly using verbal stories,
we found that no ToM regions showed greater activity for the more embedded beliefs
(Koster-Hale & Saxe, 2011). Other brain regions did show more activity—regions
involved in language processing and regions involved difficult memory and cognitive
control tasks, like dorsolateral prefrontal cortex (DLPFC). Our interpretation of these
results is that embedding beliefs inside other beliefs makes the reasoning problem
harder, but does not lead to greater ToM processing, per se. This finding converges with
results from patient populations and aging adults showing that failure to pass secondorder false belief tasks may, in fact, be due to domain-general impairment, rather than
diminished ToM processing (Slessor, Phillips, & Bull, 2007; Zaitchik, et al., 2006).
An alternative approach draws from a model put forward in other domains, such as
vision and decision making. Predictive coding posits that neural systems make forwardlooking predictions about incoming information. In this model, neural signals contain
information not about the currently perceived stimulus, but about the difference between
the observed and the predicted stimulus. There has recently been increasing interest in
the idea of predictive coding as a unifying framework for understanding neural
computations across many domains (e.g., Clark 2013). In Chapter 5, I consider this
theoretical framework in more detail, and explore its application to theory of mind.
40 of 179
A First Step
Human social life depends on understanding other people’s behavior: why they do what
they do, and what they are likely to do next. However, these actions are merely the
output of an unobservable, internal causal structure: the person’s goals and intentions,
beliefs and desires, preferences and personality traits. A cornerstone of the human
capacity for social cognition is the ability to reason about these invisible causes: having
a “theory of mind”. The literature reviewed in this chapter presents converging evidence
that reasoning about the minds of other people reliably and selectively recruits specific
cortical regions. Yet, while we have an understanding of which parts of the brain support
social inference, we have remarkably little insight into how these neural substrates
function at a computational level. A key next question in social cognitive neuroscience is
to investigate how neural populations encode these concepts.
To move past where in the brain, to how, a first step is asking whether we have the tools
and theoretical understanding to successfully examine brain regions that support social
cognition in more detail. The goal of this thesis is to demonstrate that it is possible to
use functional neuroimaging to ask questions about neural representations in social
cognition. I present evidence that neuroimaging can find abstract, behaviorally relevant
features of mental state inferences inside cortical regions that support social cognition,
by demonstrating:
(i)
that the neural code in brain regions that support social cognition contains
information which distinguishes between mental state representations (rather than
mental state representations from other types of representation), and that this
code is detectable using fMRI (Chapter 2),
(ii)
that these neural distinctions are behaviorally relevant: they reflect distinctions
that we know matter, based on cognitive and developmental work on theory of
mind, and are sensitive to individual differences in social cognition (Chapter 2),
(iii)
that these neural distinctions are robust in the face of profound differences in
developmental experience (Chapter 3),
(iv)
that these neural distinctions are part of a more complex feature space of mental
state representation, and are coded along independent, continuous dimensions
(Chapter 4),
(v)
that an existing framework of neural coding, predictive coding, can provide a
unifying framework for interpreting — and making new predictions about — the
kinds of representations we find in these brain regions.
In Chapter 2, I demonstrate that functional neuroimaging of the human brain can be
used to find neural representations of distinct features of mental states. As a first case
study, I consider intent — whether someone’s goal is aligned with the results of their
actions — when causing harm. Intent is a clear test case to look for neural
representations of mental state features. First, reasoning about intent has previously
been shown to rely on ToM brain regions: distinguishing between intentional harms and
accidents depends on the capacity for mental state reasoning and disrupting activity in
RTPJ can impair mental state reasoning for moral judgment (Young et al, 2010).
41 of 179
Second, intent has clear behavioral relevance: intentional harms are judged to be
morally worse and more punishable than accidental harms (Cushman, 2008; Baird &
Astington, 2004). And third, there are reliable individual and group differences in
reasoning about intent: individuals differ in the extent to which they weigh intent versus
outcome (Young & Saxe, 2009b; Moran et al., 2011); moreover, high-functioning
individuals with Autism Spectrum Disorders (ASD) make moral judgments based less on
intent information, compared to neurotypical (NT) participants (Grant, 2005; Moran et
al., 2011; Yang, Baillargeon, & Baillargeon, 2013). In four experiments, using multi-voxel
pattern analysis (MVPA), I find that (i) in NT adults the RTPJ shows reliable and distinct
spatial patterns of responses across voxels for intentional versus accidental harms, and
(ii) individual differences in this neural pattern predict differences in participants’ moral
judgments. These effects are specific to RTPJ. By contrast, (iii) this distinction is absent
in adults with ASD. Together, these data provide a proof of concept: fMRI can be used
to find neural representations inside ToM brain regions that reliably track differences in
behavior, both at the individual level, and the clinical level.
Having found that information about different types of mental states can be found in ToM
brain regions, we need to ask, can we find other features of mental state inference, and
can we learn anything about their format and developmental origin? In Chapter 3, I
show that ToM brain regions contain explicit, abstract representations of another feature
of others’ mental states: perceptual source. I find that in the RTPJ of sighted adults, the
pattern of neural response distinguished the source of the mental state (did the
protagonist see or hear something?). Moreover, I find that these representations persist
in the face of profound differences in developmental history (congenital blindness):
neural representations of perceptual source were preserved in congenitally blind adults,
who have no first-person experience with sight. The fact that blind people’s inferences
about how other people see rely on a neural code that parallels that of sighted people
provides a window into fundamental questions about the human capacity to think about
one another’s thoughts: we can ask both what kinds of representations people form
about others’ minds, and how much these representations depend on the observer
having had similar mental states themselves.
In Chapter 4, I take a first step towards looking for a feature space of mental state
inference, by simultaneously testing four unique features of beliefs. I find that that social
brain regions contain representations of epistemic features (why does someone believe
what they believe?) and emotional features (how do they feel based on what they
believe?) of others’ beliefs. Using MVPA, I find that the patterns of activity in RTPJ and
right STS contain complementary neural codes, which reliably distinguish stimuli in
which a person’s belief is based on reliable and unambiguous evidence, from stimuli in
which their belief based on unreliable and ambiguous evidence. Moreover, I show that
the neural code tracks item-wise differences along these abstract dimensions:
classification accuracy for each stimulus was predicted by independent ratings of the
strength of the evidence in that stimulus. I also find evidence for a representation of
evidence modality (visual vs auditory) in bilateral TPJ, and of emotional valence in
ventral MPFC. Together, these data suggest that the cortical regions implicated in
mental state inference contain neural representations of epistemic and emotional
42 of 179
features of others’ beliefs, and that these features are represented along continuous,
abstract dimensions. Finally, in Chapter 5, I take a more global view, and consider how we might understand
the activity we observe in these regions, and how we might go about generating new,
testable hypotheses. I extend a model from vision and neuroeconomics — predictive
coding — and explore its application to the neural basis of social cognition. Predictive
coding posits that neural systems make forward-looking predictions about incoming
information. Neural signals contain information not about the currently perceived
stimulus, but about the difference between the observed and the predicted stimulus. I
review evidence that, across ToM brain regions, neural responses to depictions of
human behavior, from biological motion to trait descriptions, exhibit a key signature of
predictive coding: reduced activity to predictable stimuli. I propose a series of
experiments to distinguish predictive coding from alternative explanations of this
response profile, and argue that this approach could both further our understanding of
the representations inside ToM brain regions, and our understanding of the distribution
of labor across them.
Taken together, this work provides a key next step to understanding the neural basis of
theory of mind: demonstrating that it is possible to find abstract, behaviorally relevant
features of mental state inferences inside cortical regions that support social cognition.
In Chapter 6, I discuss how these results lay the groundwork that will allow us to pursue
the research lines suggested above, and lay out some of the next steps in this research
program.
43 of 179
This page intentionally left blank
44 of 179
Chapter 2.
Decoding moral judgments from neural
representations of intentions 6
Intentional harms are typically judged to be morally worse than accidental harms.
Distinguishing between intentional harms and accidents depends on the capacity for
mental state reasoning (i.e., reasoning about beliefs and intentions), which is
supported by a group of brain regions including the right temporo-parietal junction
(RTPJ). Prior research has found that interfering with activity in RTPJ can impair
mental state reasoning for moral judgment, and that high-functioning individuals with
Autism Spectrum Disorders (ASD) make moral judgments based less on intent
information, compared to neurotypical (NT) participants. Three experiments, using
multi-voxel pattern analysis (MVPA) find that (i) in NT adults the RTPJ shows reliable
and distinct spatial patterns of responses across voxels for intentional versus
accidental harms, and (ii) individual differences in this neural pattern predict
differences in participants’ moral judgments. These effects are specific to RTPJ. By
contrast, (iii) this distinction was absent in adults with ASD. We conclude that MVPA
can detect features of mental state representations (e.g., intent), and that the
corresponding neural patterns are behaviorally and clinically relevant.
6
A version of this chapter was published as Koster-Hale, J., Saxe, R. R., Dungan, J., & Young, L. L.
(2013). Decoding moral judgments from neural representations of intentions. PNAS, 110(14), 5648–5653.
doi:10.1073/pnas.1207992110
45 of 179
Introduction
Thinking about another’s thoughts increases metabolic activity in a specific group of
brain regions. These regions, which comprise the ‘theory of mind network’, include the
medial prefrontal cortex (MPFC), precuneus (PC), right superior temporal sulcus
(RSTS), and bilateral temporal-parietal junction (TPJ). While many studies have
investigated the selectivity and domain specificity of these brain regions for theory of
mind (e.g., Aichhorn et al., 2009; Saxe & Kanwisher, 2003), a distinct but fundamental
question concerns the computational roles of these regions: which features of people’s
beliefs and intentions are represented, or made explicit, in these brain regions? Prior
work has focused on where in the brain mental state reasoning occurs, whereas the
present research builds on this work to investigate how neural populations encode
these concepts.
A powerful approach for understanding neural representation in other domains has been
to ask which features of a stimulus can be linearly decoded from a population of
neurons. For example, in the ventral visual stream (involved in object recognition), lowlevel stimulus properties like line-orientation and shading are linearly decodable from
small populations of neurons in early visual areas (e.g., V1), whereas in higher-level
regions the identity of an object becomes linearly decodable, and invariant across
viewing conditions (DiCarlo, Zoccolan, & Rust, 2012; Kamitani & Tong, 2005). These
results suggest that as information propagates through the ventral pathway, the neural
response is reformatted to make features that are relevant to object identity more
explicit to the next layer of neurons (DiCarlo et al., 2012).
A decoding approach can be similarly applied to functional magnetic resonance image
(fMRI) data, using multi-voxel pattern analysis (MVPA) to examine the spatial pattern of
neural response within a brain region. If a distinction between cognitive tasks, stimulus
categories, or stimulus features is coded in the population of neurons within a brain
region, and if the subpopulations within the region are (at least partially) organized into
spatial clusters or maps over cortex (Dehaene & Cohen, 2007; Formisano et al., 2003),
then the target distinction may be detectable in reliable spatial patterns of activity
measurable with fMRI (Haynes & Rees, 2006; Kriegeskorte, Bandettini, & Bandettini,
2007; Norman, Polyn, Detre, & Haxby, 2006). MVPA has therefore been used to identify
categories and features that are represented within a single region, (e.g., Kamitani &
Tong, 2006; Mahon & Caramazza, 2010; Peelen, Wiggett, & Downing, 2006), and to
relate these representations to behavioral performance (Haynes & Rees, 2006; Norman
et al., 2006; Raizada et al., 2010).
Compared to object recognition, much less is known about the cognitive and neural
mechanisms that support theory of mind. However, linear separability of the neural
response could serve as a diagnostic measure of the core features and local
computations even in this abstract domain. We therefore asked whether the spatial
pattern of response in theory of mind brain regions could be used to decode a feature
that has previously been shown to be critical for theory of mind: whether an action was
performed intentionally or accidentally.
46 of 179
The distinction between intentional and accidental acts is particularly salient in the case
of moral cognition. Adults typically judge the same harmful act (e.g., putting poison in a
drink, failing to help someone who is hurt, making an insensitive remark) to be more
morally wrong and more deserving of punishment when committed intentionally versus
accidentally (e.g., Cushman, 2008). These moral judgments depend on individuals’
ability to consider another person’s beliefs, intentions, and knowledge, and emerge
relatively late in childhood, around age 6-7 years (e.g., Baird & Astington, 2004).
Individuals with Autism Spectrum Disorders (ASD), who are disproportionately impaired
on tasks that require them to consider people’s beliefs and intentions (Baron-Cohen,
1995; Peterson, Wellman, & Liu, 2005), are also impaired in using information about an
innocent intention to forgive someone for accidentally causing harm (Grant, 2005;
Moran et al., 2011; Yang, Baillargeon, & Baillargeon, 2013) but see (Baez et al., 2012).
The RTPJ is particularly implicated in these moral judgments. In prior research, (i)
increased TPJ activation is related to greater consideration of mitigating intentions and
more lenient punishment (Buckholtz et al., 2008; Yamada et al., 2012), (ii) individual
differences in the forgiveness of accidental harms are correlated with the magnitude of
activity in the RTPJ at the time of the judgment (Young & Saxe, 2009b), and (iii)
interfering with activity in the RTPJ shifts moral judgments away from reliance on mental
states (Young, Camprodon, Hauser, Pascual-Leone, & Saxe, 2010a).
Given the importance of intent for moral judgments of harms, we predicted that one or
more of the brain regions in the theory of mind network would explicitly encode this
feature of others’ mental states in neurotypical (NT) adults. That is, we predicted that (i)
while participants read about a range of harmful acts, we would be able to decode
whether the harm was intentional or accidental based on the spatial pattern of activity
within theory of mind brain regions. We tested this prediction in three experiments with
NT adults. We also investigated (ii) whether the robustness of the spatial pattern within
individuals would predict those individuals’ moral judgments, and (iii) whether, in a fourth
experiment, high-functioning adults with ASD, who make atypical moral judgments of
accidental harms, would show atypical patterns of neural activity in pattern or
magnitude.
In all four experiments, participants in the scanner read short narratives in which
someone caused harm to another individual, intentionally or accidentally, as well as
narratives involving no harm. Participants in Experiments 1 and 2 made a moral
judgment about the action. Participants in Experiment 3 made true/false judgments
about facts from the narratives. In Experiment 4, high-functioning adults with ASD read
and made moral judgments about the same narratives as in Experiment 1.
47 of 179
Methods
Participants
Experiment 1: Twenty-three right-handed members of the local community (aged 18-50
years, mean=27, 7 women). Experiment 2: Sixteen right-handed college
undergraduate students (aged 18-25 years, 8 women). Experiment 3: Fourteen righthanded college undergraduate students (aged 18-25 years, 8 women). Experiment 4:
Sixteen individuals diagnosed with Autism Spectrum Disorders (aged 20-46 years,
mean=31, 2 women). Participants in Exp.4 were recruited via advertisements placed
with the Asperger’s Association of New England. All participants were prescreened
using the Autism Quotient questionnaire (AQ; Baron-Cohen, 2001). ASD participants
(mean=32.6) scored significantly higher on the AQ than the NT participants from Exp.1
(mean=17.3; t(25.5)=6.4, p<0.0001).
ASD participants underwent both the Autism Diagnostic Observation Schedule (Joseph,
Tager-Flusberg, & Lord, 2002; ADOS; Lord et al., 2000) and impression by a clinician
trained in both ADOS administration and diagnosis of ASD. All ASD participants
received a diagnosis of an Autism Spectrum Disorder based on their social ADOS score
(criterion of ≥ 4; mean=6.4), communication ADOS score (criterion of ≥ 2; mean=3.1),
and total ADOS score (criterion ≥ 7; mean=9.5), and on clinical impression based upon
the diagnostic criteria of the DSM-IV (American Psychiatric Association & American
Psychiatric Association Task Force on DSM-IV, 2000). The NT (Exp.1) and ASD (Exp.4)
groups did not differ in age (NT mean=27, ASD mean=31, t(35.26)=1.32, p=0.2) or IQ
(NT mean=121; ASD=120; t(28.3)=0.26, p=0.8). Additionally, all analyses were run with
one-to-one matched subsets of participants (both n=15, see Appendix).
All participants were native English speakers, had normal or corrected-to-normal vision,
gave written informed consent in accordance with the requirements of Institutional
Review Board at MIT, and received payment. Data from Exp.2 and 3 have previously
been published in papers analyzing the magnitude but not the pattern of response in
each region (Young & Saxe, 2008; 2009a).
fMRI Protocol and Task
Experiments 1 and 4
Participants were scanned while reading 60 stories (Figures 2.1): 12 described harm
caused intentionally (e.g. knowingly kicking someone in the face), 12 described harm
caused accidentally (e.g. kicking without seeing the person), 12 described neutral
actions (e.g. eating lunch), and 24 stories describing disgusting but not harmful actions
(e.g. smearing feces on one’s own face, not analyzed here). Stories were presented in 4
cumulative segments, describing the background (6s), action (+4s), outcome (+4s) and
intention (+4s). After each story, participants made a moral judgment of the action (“How
much blame should you get?”) from “none at all” (1) to “very much” (4), using a button
press. Behavioral data from 5 ASD and 3 NT participants were lost due to error (n=4) or
to theft of experimental equipment (n=4). Ten stories were presented in each 5.5 min
run; the total experiment, six runs, lasted 33.2 min.
48 of 179
Experiment*1+4
Accidental
Intentional
Experiment*2
Accidental*Harm
Intentional*Harm
Background
Action
Outcome
Intent
You%had%absolutely*no*idea*
Your%family%is%over% You%grind%up% Your%cousin,%
about%your%cousin's%allergy%
for%dinner.%You%wish%
some%
one%of%your%
when%you%added%the%peanuts.
to%show%off%your%
peanuts,%add%
dinner%
culinary%skills.%For% them%to%that% guests,%is%
one%of%the%dishes,%
dish,%and%
severely%
You%knew%about%your%cousin's%
adding%peanuts%will%
serve%
allergic%to%
peanut%allergy%when%you%added%
really%bring%out%the%
everyone.
peanuts.
the%peanuts%to%the%dish.
flavor.
Background
Foreshadow
Belief
The%container%is%
labeled%“sugar”,%
Grace%and%her%friend%
so%Grace%believes%
are%taking%a%tour%of%a%
The%white%
that%the%white%
chemical%plant.%When%
powder%is%a%very% powder%is%regular*
Grace%goes%over%to%the%
toxic%substance%
sugar.
coffee%machine%to%pour%
left%behind%by%a%
some%coffee,%Grace’s%
The%container%is%
scientist,%and%
friend%asks%for%some%
labeled%“toxic”,%
deadly%when%
sugar%in%hers.%There%is%
so%Grace*believes%
ingested.
white%powder%in%a%
that%the%white%
container%by%the%coffee.
powder%is*a*toxic*
Outcome
Grace%puts%the%
substance%in%her%
friend’s%coffee.%
Her%friend%drinks%
the%coffee%and%
dies.%
substance.*
Experiment*3
Accidental*Harm
Intentional*Harm
Background
Belief
Outcome
Peter%believes%that%it%is*safe%to%
wade%in%the%pond%because*other*
Peter%is%traveling%in%
Africa%with%a%friend.%His% tourists*around*them*are*doing* Malarial%mosquitoes%
it*too*and*seem*to*be*enjoying*
friend%sees%a%pond%and%
actually%live%in%the%
themselves.*
wants%to%go%wading%in%it%
pond.%A%single%bite%is%
because%it%is%very%hot.%
Peter%believes%that%it%is*unsafe*to% enough%to%create%an%
Peter%encourages%his%
wade%in%the%pond*because%Africa* infection,%so%the%pond%
friend%to%wade%in%the%
is%unsafe%to%wade%in.
is*known*for*malarial*
pond.%
mosquitoes,*and*mosquitoes*
congregate*around*water.*
Figure 2.1. Sample stimuli from Exp. 1 and 4 (top), Exp. 2 (middle), and Exp. 3 (bottom)
Participants were scanned while reading 60 stories (Figure 2.1). Stories were
presented in the second person, using present tense. Each participant read 12 stories
describing someone harming another individual caused accidentally, 12 describing
someone harming another individual intentionally, 12 neutral actions (described below),
and 24 stories describing disgusting but not harmful actions (e.g., consensual incest,
drinking blood, smearing feces on one’s own face, not analyzed here).
Each story was displayed in four cumulative segments (i.e. earlier segments remained
on the screen as later segments were added). Cumulative segments partially reduce the
49 of 179
variability in reading time across participants and items, by slowing down faster
participants and shorter items, without cutting off reading early too often in longer
segments or for slower readers.
For accidental and intentional harms, the initial three segments of the stories were
identical: information about the Background (6s), Action (+4s), and harmful Outcome
(+4s). Harmful outcomes included both physical harms (kicking someone in the face,
feeding them an allergen) and psychological harms (insulting or humiliating someone).
The only information that distinguished between these two conditions was the final
sentence, describing the Intent (+4s), which was innocent in the accidental harm
condition (e.g. “you did not see”, “you did not know” that the act would cause harm) and
negative in the intentional harm condition (e.g. “you saw”, “you realized” that the act
would cause harm). The accidental and intentional sentences were matched pairwise by
changing the fewest possible words in the Intent segment, thus preserving length (see
Figure 2.1 for examples; mean=14 words, standard deviation=1.9, range=9-18). Across
subjects, each scenario (i.e. Background+Action+Outcome) was equally likely to occur
in the Intentional or Accidental condition. Each participants saw either the intentional or
the accidental version of each scenario, but not both.
For the neutral actions, a separate set of scenarios described a neutral Background and
Action (e.g. giving a class presentation, eating lunch). The Outcome was mildly positive
(the teacher liked the graph) or negative (a little spilled ketchup); and the Intent
described either knowledge or no knowledge of the outcome (e.g. “you did/did not
realize” that you spilled ketchup). Because there were only 12 neutral stories in total, we
used the averaged the response for all neutral stories in the analyses.
In the scanner, after each story, participants made a moral judgment of the action (“How
much blame should you get?”) from “none at all” (1) to “very much” (4), using a button
press. Participants were given 4 seconds to respond before the screen was blanked.
Thus, the whole trial lasted 22 seconds. A 10 s rest block (blank screen) occurred after
each trial.
Stories were presented in a pseudorandom order, such that no condition was
immediately repeated, each condition was equally likely to occur in each position in the
run, and each condition was equally likely to follow each other condition. Ten stories
were presented in each 5.5 min run; the total experiment, six runs, lasted 33.2 min.
Stories were projected onto a screen via Matlab 5.0 running on an Apple MacBook Pro
in 40-point white font.
Experiment 2
Young et al.(Young & Saxe, 2008) report results from 17 participants; data from one
participant have been lost, leaving 16 participants.
Participants were scanned while reading 48 stories. Stories were presented in the third
person, using past tense. Each participant read stories describing 12 intentional harms,
12 accidental harms, 12 failed attempts to harm, and 12 neutral actions (Figure 2.1; for
the complete list see Young & Saxe, 2008). The focus of the current manuscript is the
accidental and intentional harm stories. All harms were physical harms, resulting in
someone’s death or serious injury.
50 of 179
Stories were presented in four cumulative segments: Background (6s); Foreshadow
(+6s); Belief (+6s); and Outcome (+6s). Half of the stories in each run were presented
with Foreshadow before Intent; the other half were presented with Intent before
Foreshadow. The Background, Foreshadow, and Outcome segments of intentional and
accidental harms were identical. The Belief segments described the content of the
character’s belief (e.g. “Leah believes that the child is about to dive into shallow water
and break his neck”) and the reason for that belief (e.g. “Because of a warning sign at
the side of the pool”). The difference between conditions was created by manipulating
the content of the belief. In the context of the foreshadow (“The child is about to dive
into the shallow end and smack his head very hard”) the action (“Leah walks by, without
saying anything to the child”), and the negative outcome, these beliefs determined
whether the harm (“The child dives in and breaks his neck”) was caused intentionally or
accidentally.
After each story, participants made moral judgments (“How morally permissible was X’s
action?”) of the action on a 3-point scale, from “forbidden (1) to “permissible” (3), using
a button press. Participants were given 4 seconds to respond before the screen was
blanked. All trials were followed by a 14s rest block. Stories were projected onto a
screen via Matlab 5.0 running on an Apple G4 laptop in 24-point white font. Behavioral
data were available for 10 participants. The methods and materials from Experiment 2
have been published in (Young & Saxe, 2008).
Experiment 3
Participants were scanned while reading 48 stories. Stories were presented in the third
person, using present tense. Each participant read stories describing 8 intentional
harms, 8 accidental harms, 8 failed attempts to harm, 8 neutral actions, and 16 actions
for which the outcome was unknown (a morally irrelevant fact was presented in place of
the outcome; Figure 2.1 and Young & Saxe, 2009a Experiment 2). Stories were
presented in cumulative segments: Background (6s), Belief (+6s), and Fact (+6s). The
focus of the current manuscript is the accidental and intentional harm stories; the
Background and Fact segments were identical for these conditions. Belief segments
described the content of the character’s belief (e.g. “Steve believes the ground beef is
safe to eat”) and the reason for that belief (e.g. “because of the expiration date.”). The
difference between conditions was created by manipulating the content of the belief. In
the context of the fact (e.g. “The meat has some invisible but deadly bacteria”) and the
action (Steve “makes a large meatloaf for all the children”), these beliefs led to either
accidental or intentional harm, although the harmful outcome itself (i.e. the children
becoming ill) was only implied, not explicit stated. After each story, participants
answered a true/false question about the Fact segment (e.g. “The meat is unsafe for
consumption: True / False.”). Participants were given 4 seconds to respond before the
screen was blanked. All stories were followed by a 14s rest block. Stories were
projected onto a screen via Matlab 5.0 running on an Apple G4 laptop in 24-point white
font. The methods and materials from Experiment 3 have been published in (Young &
Saxe, 2009a).
51 of 179
Theory of Mind Localizer task
All subjects participated in a theory of mind localizer task, contrasting stories requiring
inferences about mental state representations (e.g., thoughts, beliefs) versus physical
representations (e.g., maps, signs, photographs). Mental state stimuli versus physical
representation stimuli were similar in their meta-representational and logical complexity;
the key difference was that for the mental state stories the reader had to build a
representation of someone else’s mental state. See (Dodell-Feder, Koster-Hale, Bedny,
& Saxe, 2011; Saxe & Kanwisher, 2003) for further discussion. Stimuli and experimental
design are available at http://saxelab.mit.edu/superloc.php. This experiment was
modeled treating each trial as a block, with a boxcar lasting 14 seconds, the whole
period from the initial presentation of the story to the end of the question presentation.
Beta values were estimated, in each voxel, for stories describing mental states (“Belief”)
or physical representations (“Photo”). Then a simple contrast map was produced in
each subject, identifying voxels responding more to Belief than Photo stories, at
p<0.001 uncorrected for multiple comparisons.
Acquisition and Preprocessing
In all four experiments, fMRI data were collected in a 3T Siemens scanner at the
Athinoula A. Martinos Imaging Center at the McGovern Institute for Brain Research at
MIT, using a 12-channel head coil. Using standard echoplanar imaging procedures,
blood oxygen level dependent (BOLD) signal was acquired in 26 near axial slices using
3x3x4 mm voxels (TR=2 s, TE=40 ms, flip angle=90°). To allow for steady state
magnetization, the first 4 seconds of each run were excluded. Data processing and
analysis were performed using SPM8 (Exp.1 and 4; (http://www.fil.ion.ucl.ac.uk/spm),
SPM2 (Exp.2 and 3) and custom software. The data were motion corrected, realigned,
normalized onto a common brain space (Montreal Neurological Institute (MNI)
template), spatially smoothed using a Gaussian filter (full-width half-maximum 5 mm
kernel) and high-pass filtered (128 Hz).
Motion and Artifact Analysis
For the NT (Exp.1) and ASD (Exp.4) participants, we calculated the total motion per run
(sum of translation in three dimensions), and the number of timepoints that either (a)
deviated from the global mean signal by more than 3SD, or (b) included TR-to-TR
motion of more than 2mm.
fMRI Analysis
All fMRI data were modeled using a boxcar regressor, convolved with a standard
hemodynamic response function (HRF). The general linear model was used to analyze
the BOLD data from each subject as a function of condition. The model included
nuisance covariates for run effects, global mean signal, and an intercept term. A slow
event-related design was used. An event was defined as a single story, beginning with
the onset of text on screen and ending after the response prompt was removed.
52 of 179
Functional Localizer: Individual subject ROIs
Functional regions of interest (ROIs) were defined in right and left temporo-parietal
junction (RTPJ, LTPJ), medial precuneus (PC), and dorsal medial prefrontal cortex
(DMPFC). For each participant, we found the peak voxel in the contrast image for each
region. ROIs were then defined as all voxels within a 9mm radius of the peak voxel that
passed threshold in the contrast image (Belief > Photo, p<0.001 uncorrected, k>10);
Within-ROI Magnitude Analysis
We measured the response to each condition in each ROI. Baseline response in each
region was the average BOLD response, in that region, at at all timepoints when there
was no stimulus on the screen, excluding the first 4 seconds after the off-set of each
stimulus (to allow the hemodynamic response to decay). The percent signal change
(PSC) relative to baseline was calculated for each time point in each condition,
averaging across all voxels in the ROI and across all blocks in the condition, where
PSC(at time t)=100 × ((average BOLD magnitude for condition (at time t) – average
BOLD magnitude for fixation) / average BOLD magnitude for fixation). We averaged the
PSC across the entire presentation (offset 6s from presentation time to account for
hemodynamic lag) to estimate a single PSC for each condition, in each ROI, in each
participant (Poldrack, 2006). Using custom software to visualize the percent signal
change in a region, we extract the full timecourse in the region (i.e. set of voxels), and
then create an event-related average, by aligning and averaging the raw BOLD
response at each time point after the trial onset, per condition (rather than extracting the
beta value for the whole block). Other software packages (e.g. SPM) calculate the
baseline response as the mean over the run, which is undesirable because it will overestimate the response at baseline in a region with a strong positive response to most
blocks, and will under-estimate the response at baseline in a region that deactivates
during most blocks.
In addition, because the stimuli in the current experiment were presented cumulatively,
for Experiments 1 & 4 we separately analyzed the PSC for two phases of the trial: (1)
the initial phase (Background + Action + Outcome, first 14 s), in which intentional and
accidental harms could not be distinguished, and (2) the final phase (Intent + Decision,
final 8 s), after intentional and accidental harms were distinguished.
Within-ROI Pattern Correlation Analysis
In all four experiments, we conducted within-ROI pattern analyses. Following Haxby et
al. (2001), each participant's data were divided into even and odd runs (‘partitions’) and
the mean response (beta value) of every voxel in the ROI was calculated for each
condition. The pattern of activity was defined as the vector of beta values across voxels
within the ROI. To calculate the within-condition correlation, the pattern in one (e.g.,
even) partition was compared to the pattern for the same condition in the opposite (e.g.,
odd) partition; to calculate the across-condition correlation, the pattern was compared to
the opposite condition, across partitions. For each individual, an index of classification
was calculated as the z-scored within-condition correlation minus the z-scored acrosscondition correlation. A region successfully classified a difference in conditions if, across
individuals, the within-condition correlation was higher than the across-condition
53 of 179
correlation, using a Student's T complementary cumulative distribution function. Note
that in this procedure (unlike other machine learning approaches, e.g., support vector
machines) differences in the spatial patterns across conditions are independent of
differences in the average magnitude of response. This procedure implements a simple
linear decoder, which is, while in principle less flexible and less powerful than non-linear
decoding, preferable both theoretically and empirically (e.g., DiCarlo et al., 2012;
Kamitani & Tong, 2005; Norman et al., 2006).
Whole Brain Pattern Correlation Analysis (Searchlight)
Searchlight. In this analysis, rather than using a predefined ROI, a Gaussian kernel
(14mm FWHM, corresponding approximately to the size of the functional ROIs) is
moved iteratively across the brain. Using the same logic as the ROI-based MVPA, we
compute the spatial correlation, in each kernel, of the neural response (i.e. betas) within
conditions and across conditions; we transform the correlations using Fisher’s Z, and
subtract the across-condition from the within-condition correlation to create an index of
classification. Thus, for each voxel, we obtain an index of how well the spatial pattern of
response in the local region (i.e. the area centered on that voxel) can distinguish
between the two conditions of interest. The use of a Gaussian kernel smoothly deemphasizes the influence of voxels at increasing distances from the reference voxel
(Fedorenko, Nieto-Castañón, & Kanwisher, 2012). We created whole brain maps of the
index of classification for each subject. These individual correlation maps were
subjected to a second-level analysis using a one-sample t-test (thresholded at p<0.001,
voxelwise, uncorrected). Note that because each voxel’s discrimination index reflects
the spatial pattern in the surrounding region, the map of spatial discrimination is highly
smooth, reducing the effective number of comparisons. "
This classification procedure implements a simple linear decoder. Linear decoding,
while in principle less flexible and less powerful than non-linear decoding, is preferable
both theoretically and empirically. A non-linear classifier can decode nearly any arbitrary
feature contained implicitly within an ROI, reflecting properties of the pattern analysis
algorithm rather than the brain, which makes successful classification largely
uninformative (Cox & Savoy, 2003; DiCarlo & Cox, 2007; Goris & Op de Beeck, 2009;
Kamitani & Tong, 2005; Norman et al., 2006). Moreover, linear codes have been argued
to be more neurally plausible, reflecting the information made available to the next layer
of neurons (Bialek, Rieke, Van Steveninck, & Warland, 1991; Butts et al., 2007; DiCarlo
& Cox, 2007; Naselaris, Kay, Nishimoto, & Gallant, 2011; Rolls & Treves, 2011).
54 of 179
Results
fMRI task: Behavioral Results
Because participants used the scales in different ways (e.g., some using largely 2-4,
others using largely 1-3), we z-scored the behavioral data (see Appendix for
untransformed data).
Experiment 1: Intentional harms (1.2±0.03) were rated as more blameworthy than
accidental harms (-0.38±0.04; t(19)=-25, p<0.001), and both were rated more
blameworthy than neutral acts (-0.8±0.03; t(19)=50, p<0.001; t(19)=6.9, p<0.001).
Experiment 2: Replicating the results in Exp.1, participants judged intentional harms
(1.0±.05) to be worse than accidental harms (-0.53±.11; t(9)=16.0, p<0.001).
Experiment 3: Participants in Exp.3 did not make moral judgments of the scenarios, but
instead answered true/false questions about the content of the fact. Reaction time (RT)
was analyzed using a one way, repeated measures ANOVA, revealing a main effect of
condition, (F(1,13)=4.56, p<0.02). Post hoc t-tests revealed that RTs for answering
factual questions following non-harm stories (mean RT=2.4 sec) were not different from
RTs following accidental harms (mean 2.3 sec; t(13)=1.65 p=0.12), or intentional harms
(mean 2.5 s, t(13)=1.49 p=0.16), but that responses following intentional harms were
slower than responses following accidental harms (t(13)=2.87, p< 0.01).
Experiment 4: When making moral judgments, ASD participants, like NT participants
from Exp.1, judged intentional harms (1.0±0.09) more blameworthy than accidental
harms (-0.23±0.08, t(10)=8.0 p<0.0001), and both intentional and accidental harms
were rated worse than neutral acts (-0.77±0.08, t(10)=11.5, p<0.0001; t(10)=4.3,
p=0.0015).
Group Comparison: A mixed effects ANOVA crossing Group (NT in Exp.1, ASD in Exp.
4) by Condition (accidental, intentional, z-scored ratings) yielded a main effect of
Condition (F(1,29)=446.9, p<0.0001) and a Group by Condition interaction (F(1,29)=4.7,
p=0.03). A planned comparison t-test (Moran et al., 2011) revealed that ASD adults
assigned more blame for accidental harms than NT adults (t(29)=1.9, p=0.03, onetailed). An additional post-hoc t-test revealed that ASD adults also assigned less blame
to intentional actions (t(29)=2.1, p=0.04).
Motion and Artifact Analysis Results
There was no difference in total motion between NT (Exp.1, mean=0.24mm/run) and
ASD participants (mean=0.23mm/run, t(37)=0.24, p=0.81), or in the number of outliers
per run (NT: 3.7±.86; ASD: 3.4±1.2, t(37)=0.2, p=0.8).
fMRI Results: Functional Localizer
Replicating studies using a similar functional localizer task (e.g., Saxe & Kanwisher,
2003), we localized four theory of mind brain regions showing greater activation for
mental state stories (e.g., describing false beliefs) compared to physical state stories
(e.g., describing “false” or outdated physical representations, such as photographs;
p<0.001, k>10) in the majority of individual participants. Four ROIs were identified for
55 of 179
each subject, based on the contrast “belief > photo” thresholded at P < 0.001. For
experiments 1–3 (NT), these were as follows: RTPJ (52/53), LTPJ (51/53), pre- cuneus
(PC; 51/53), and dorsal medial prefrontal cortex (DMPFC; 43/53). For experiment 4
(ASD), these were as follows: RTPJ (16/16), LTPJ (15/16), PC (16/16), and DMPFC
(7/16) Table 2.1 reports the average position of the peak and the average and SD of the
ROI size for each ROI for each experiment. All subsequent analyses are conducted
using individually-defined regions of interest (ROIs).
Experiment*1
Region
size
Experiment*2
Average-PeakVoxel
x
y
z
size
Experiment*3
Average-PeakVoxel
x
y
z
size
Experiment*4*(ASD)
Average-PeakVoxel
x
y
z
size
Average-PeakVoxel
x
y
z
RTPJ
260'
(17)
54
G54
22
151'
(18)
56
G56
22
106'
(22)
53
G55
21
194'
(24)
55
G55
24
LTPJ
144'
(26)
G51
G56
22
121'
(21)
G50
G63
26
178'
(23)
G54
G58
19
156'
(35)
G51
G60
26
PC
253'
(25)
0
G58
35
183'
(25)
G1
G58
39
192'
(20)
1
G58
37
260'
(23)
1
G54
33
DMPFC
142'
(23)
3
44
24 87'(17)
G2
58
29
130'
(19)
2
54
36
175'
(55)
6
54
31
Table 2.1: Individual subject ROIs. Individual subject ROIs are defined as all voxels within a 9mm sphere
around the peak voxel (T of mental-physical from the localizer) identified in each region. Size: average
number ± standard deviation of voxels included in individual ROIs. Average Peak Voxel: average
coordinates of the peak voxels, across participants, reported in Montreal Neurological Institute (MNI)
space.
fMRI Results: Response Magnitude
Experiment 1 (NT): Averaged over the whole trial, harmful actions elicited a higher
response than neutral acts in all four ROIs (RTPJ, LTPJ, PC, DMPFC; all t>4.6,
p<0.0003).
In the initial phase only (Background, Action, Outcome), we compared the response to
harms versus neutral acts. In all four ROIs, harmful actions elicited a higher response
than neutral acts (all t>4, p<0.001).
In the final 8 s of the trial, after the intention had been revealed, the response in the
RTPJ of NT adults was higher for accidental than intentional harms (mean percent
signal change from rest, accidental: 0.1±0.04, intentional: 0.01±0.04, t(21)=3.59
p=0.002). LTPJ showed a similar trend (t(21)=1.82 p=0.08); there was no difference in
the other two regions. There was no difference in PC or DMPFC (both t<0.2, p>0.3).
Experiment 4 (ASD): Averaged over the whole trial, harmful actions elicited a higher
response than neutral acts in all four ROIs (RTPJ, LTPJ, PC, DMPFC; all t>3, p<0.02).
In the initial phase only (Background, Action, Outcome), we compared the response to
harms versus neutral acts. In all four ROIs, harmful actions elicited a higher response
than neutral acts (all t>3.5, p<0.0012). In the final phase only, after the intention had
56 of 179
been revealed, in ASD adults, no region showed a significant difference between
accidental and intentional harms (all t<1, p>0.3).
Group Comparison: We directly compared the magnitude of responses across groups
(ASD, NT) in three ways. First we compared the response to all harms vs all neutral
acts, averaged across the whole trial. No region showed a main effect of Group or a
Group by Condition interaction (all F<2.6, all p>0.12).
RTPJ
0.6
Harm
Accidental
Percent Signal Change
Intentional
0.4
NT
ASD
0.2
0.0
intent information
−0.2
0
5
10
Seconds
15
20
25
Figure 2.2: Percent Signal Change in the RTPJ of ASD and NT adults, for accidental harms,
intentional harms, and neutral stories. NT: n=22, ASD: n=16. Error bars indicate within-participant
SEM. In the initial phase (Background, Action, and Outcome, seconds 0-14, plus 4 s for hemodynamic
lag), both NT and ASD show a significant increase in mean activity in RTPJ for harms stories, relative to
rest. After information about intent was revealed (second 14 plus 4 s for hemodynamic lag), NT showed
increased activity for accidental harms, relative to intentional harms, while ASD showed no difference.
Second, we compared the response to all harms vs all neutral acts during just the initial
phase (Background, Action, and Outcome) of each trial. In PC, the response trended
towards being overall higher in the ASD group; an ANOVA crossing Group (ASD, NT) by
Condition (harms, neutral) revealed a marginal main effect of group (F(1,35)=3.37,
p=0.08). There were no other main effects of Group or any Group by Condition
interactions (all F<1.7, p>0.2)."
Finally, we compared the response to accidental vs intentional harms, specifically during
the final phase of the trial. A Group (NT, ASD) by Condition (accidental, intentional)
ANOVA, revealed a Group by Condition interaction in RTPJ (F(1,36)=5.37, p=0.03),
such that NT showed a difference between accidental and intentional harms, while ASD
did not). No other effects of Group, or Group by Condition interactions, were significant
(all F<1.3, p>0.25). See Figure 2.2 for the Percent Signal Change for both groups.
57 of 179
Experiment 2 & 3: See (Young & Saxe, 2008; 2009a) for analysis of the response
magnitude in Experiments 2 and 3.
fMRI Results: Pattern Analysis for Neurotypical
Within ROI Pattern Analysis for Experiment 1
Harm vs. Neutral: Multi-voxel pattern analyses revealed reliably distinct patterns of
neural activity for harmful (intentional, accidental) versus neutral acts — the pattern
generated by stories in one category (i.e., harmful or neutral) was more correlated with
the pattern from other stories in the same category than in the opposite category — in
three of four ROIs: RTPJ, LTPJ, and PC (RTPJ: within=1.3±0.10, across=1.1±0.10,
t(21)=3.2, p=0.002; LTPJ: within=1.6±0.05, across=1.4±0.06, t(21)=3.3, p=0.002; PC:
within=1.1±0.09, across=0.86±0.1, t(21)=3.4, p=0.001). DMPFC showed the same trend
(DMPFC: within=1.2±0.09, across=1±0.09, t(17)=1.7, p=0.056; Figure 2.3 A). All
correlations are Fisher Z transformed to allow statistical comparisons with parametric
tests; see Appendix for an analysis of untransformed data.
C
NT
Within
Across
ASD
RTPJ
Within
Across
2
.5 1 1.5
0
2
Exp. 4 ASD
.5 1 1.5
Exp.1 NT
∗
Exp. 1 NT
Exp. 4 ASD
∗
∗
Exp. 2 NT
Exp. 3 NT
0
∗
correlation (Z)
∗
correlation (Z)
2
.5 1 1.5
B Accidental vs Intentional
0
correlation (Z)
A Moral vs Neutral
Figure 2.3: MVPA results from Experiment 1-4. (A) Neurotypical (n=23) and ASD adults (n=16) show
pattern discrimination in the RTPJ for moral vs. neutral actions, with higher within-condition correlations
than across- condition correlations. (B) NT adults show pattern discrimination for accidental vs. intentional
harms in RTPJ, a finding replicated across Experiments 2 and 3 (n=16,14), but adults with ASD do not
show discrimination of intent. High overall correlation in ASD, combined with matched motion parameters
for both groups, suggest that this is not due to noise, but rather a stereotyped response across both
accidental and intentional harms in adults with ASD. (C) The RTPJ in a single participant. Error bars
indicated SEM.
58 of 179
NT
ASD
W
Accidental vs. Intentional: Only in RTPJ did the pattern of activity distinguish between
accidental and intentional harms (within=1.2±0.12, across=1.1±0.12, t(21)=2.2, p=0.02).
No other regions showed sensitivity to intent (all correlation differences<0.02, all p>0.3,
Figure 2.3 B).
Within ROI Pattern Analysis for Experiment 2 and 3
Exp.2 and 3 replicate Exp.1. In only the RTPJ did MVPA reveal reliably distinct neural
patterns for intentional and accidental harms (Exp.2 RTPJ: within=1.1±0.13,
across=0.91±0.10, t(15)=2.6, p=0.01; Exp.3 RTPJ: within=0.42±0.10, across =
0.24±0.10, t(13)=2, p=0.034; all other regions: correlation differences<0.1, p>0.1,
Figure 2.3 B).
RTPJ
LTPJ
PC
MPFC
C
Moral Judgment
1 1.5 2
.5
B
Pattern Correlation
0 .5 1 1.5
A
∗∗
Within
Across
∗∗
r2=.28
Within
Across
∗
n = 29
r2=.14
Within
Across
ns
n = 29
r2=.04
Within
Across
ns
n = 29
-0.7 -0.3 1 0.5 0.9
-0.6 -0.3 0 0.3 0.6
-0.8 -0.4 0 0.4 0.8
Pattern Discrimination
Pattern Discrimination
Pattern Discrimination
r2=.02
-0.7 -0.4
n = 26
0
0.4 0.7
Pattern Discrimination
Figure 2.4: MVPA results from Experiments 1, 2 & 3. (A) theory of mind regions from a single
participant. (B) Across three experiments, neurotypical adults (n=53) show pattern discrimination for
accidental vs. intentional harms in the right TPJ, but not in any other theory of mind region. (C) Moreover,
individual differences in pattern discrimination (within condition correlation minus across condition
correlation) predict differences in behavioral moral judgment (ratings of intentional minus accidental) in
NT adults in RTPJ and LTPJ (n=29, Exp. 1&2), such that adults with more discriminable patterns are
more forgiving of accidental harms relative to intentional harms. Error bars indicated SEM.
Pooling the data across all three experiments increased our power to detect results in
neural regions beyond RTPJ. Again, MVPA revealed distinct patterns for accidental and
intentional harms in RTPJ (within=0.96±0.08, across=0.81±0.08, t(51)=3.9, p=0.0002)
and no other region (all differences<0.1, p>0.1, Figure 2.4 B).
59 of 179
Behavioral and Neural Correlation
In Exp.1 and 2, NT participants provided moral judgments of each scenario in the
scanner, allowing us to determine whether behavioral responses were related to neural
pattern. In both experiments, we found that only in the RTPJ, the difference between
intentional and accidental harms in individuals’ moral judgments was correlated with the
neural classification index (within-across condition correlation; Exp.1: r(17)=0.6;
p=0.007; Exp.2: r(8)=0.65; p=0.04). The correlation between neural pattern and
behavior was also significant after combining the data from both experiments (r(27)=0.6,
p=0.0005; Figure 2.4 C). Though neither Exp.1 nor Exp.2 showed a significant
correlation in LTPJ alone, combining data from both revealed a significant correlation in
LTPJ as well (r(27)=0.38, p=0.04). The other regions showed no significant correlation
with behavioral judgments (all r<0.4, p>0.1).
Whole Brain Searchlight Analysis for Experiment 1,2,3 (NT)
Combining across all NT participants (n=53, Exp 1, 2, 3) we found only one small region
that discriminated (p<0.001, voxel-wise, uncorrected) between accidental and
intentional harms in the fusiform gyrus (peak voxel MNI coordinates: [30,-54,-14],
Figure 2.5 (top): Whole Brain Searchlight Analysis in Neurotypical Adults for Accidental vs.
Intentional Harms. (n=53, p<0.001, uncorrected) Sagittal slice, showing region in left fusiform gyrus,
and, and axial slices 40- 110.
Figure 2.5 (bottom): Whole Brain Searchlight Analysis in Behaviorally Sensitive Neurotypical
Adults for Accidental vs. Intentional Harms. (n=15, p<0.001, uncorrected) Surface, showing region in
anterior RTPJ, and slices 40-110. Participants were median-split based on the difference in their moral
judgments between intentional and accidental harms (Exp.1&2). RTPJ is the only region that
discriminated intentional and accidental harms in the group with the largest difference in their ratings of
accidental and intentional harms.
60 of 179
Figure 2.5). We then median-split participants based on the difference in their moral
judgments between intentional and accidental harms (Exp.1&2). In participants showing
a larger behavioral effect (n=15) the only region that discriminated between intentional
and accidental harms was in RTPJ (peak voxel MNI coordinates: [52, -60, 28], Figure
2.5). In the remaining participants (n=15), no significant voxels were found.
fMRI Results: Pattern Analysis for ASD
Within ROI Pattern Analysis for Experiment 4 (ASD)
Harm vs. Neutral: As in NT controls, pattern analyses revealed a separation in the
pattern of response for harmful versus neutral acts in ASD. Using individual ROIs,
significant discrimination was found in RTPJ, LTPJ, and PC (RTPJ: within=1.4±0.13,
across=1.2±0.16, t(15)=4.1, p=0.0005; LTPJ: within=1.5±0.10, across=1.3±0.16,
t(14)=2.1, p=0.027; PC: within=1.2±0.14, across=1.1±0.14, t(14)=1.9, p=0.04, Figure
2.3). The effect was non-significant in DMPFC (within=1.2±0.11, across=1.2±0.12,
t(6)=0.21, p=0.42), possibly because an ROI for DMPFC was defined in only 7/16
participants.
Accidental vs. Intentional: Accidental vs. Intentional: Accidental and intentional harms
did not elicit distinct patterns in any theory of mind ROI in ASD: RTPJ: within=1.3±0.11,
across=1.3±0.12, t(15)=1.1, p=0.86; LTPJ: within=1.4±0.12, across=1.4±0.14, t(14)=1.1,
p=0.86; PC: within=1.2±0.14, across=1.2±0.15, t(14)=1, p=0.16; DMPFC:
within=1±0.15, across=1.1±0.08±, t(6)=0.59, p=0.71, Figure 2.3).
Behavioral and Neural Correlation (ASD)
There was a marginal negative correlation between participants’ moral judgments of
accidental vs intentional harms and pattern discrimination in RTPJ (r(9)=-.54, p=.09).
Note that this effect is in the reverse direction of the correlation observed in NT adults.
There was no correlation with behavior in any other region (all r<0.2, p>0.5). There was
also no correlation with symptom severity (ADOS score) in any region.
Whole Brain Searchlight for Experiment 4 (ASD)
No regions discriminated between accidental and intentional harms (p<0.001, voxelwise, uncorrected).
ASD symptom severity in Experiment 4 (ASD)
Some prior work has found correlations between symptom severity and neural
information, either in mean signal (e.g., Wang, 2006) or in pattern discriminability (e.g.,
Coutanche, Thompson-Schill, & Schultz, 2011). However, in this data set, we found no
significant correlation between neural pattern and ADOS symptom severity scores(Lord
et al., 2000)in any region (RTPJ: r(16)=0.33, p=0.22; LTPJ: r(15)=0.05, p=0.85; PC:
r(15)=.14, p=.61; DMPFC: r(7)=0.04, p=0.93).
Group Comparison: NT vs ASD
Within ROI Pattern Analysis
61 of 179
Moral vs. Neutral: A Group (ASD, NT) x Pattern (within, across) ANOVA revealed that
NT and ASD participants show strong, and equally robust, neural discrimination in
response to moral violations versus neutral actions in their RTPJ, LTPJ, and PC, with a
main effect of Pattern (within > across), no effect of Group, and no interaction in all
three ROIs (RTPJ: Pattern (F(1,36)=25.9, p<0.0001), Group (F(1,36)=0.5, p=0.5),
interaction (F(1,36)=0.9, p=0.3); LTPJ: Pattern (F(1,35)=13.3, p = 0.0008), Group
(F(1,35) = 1.2, p = 0.3), interaction F(1,35)=0.3, p=0.6; PC: Pattern, (F(1,35)=15.1,
p=0.0004), Group (F(1,35)=1.7, p=0.2), interaction (F(1, 35) =1.18, p=0.3)). There were
no significant effects in DMPFC (Pattern (F(1,23)=2.7, p=0.1), Group, (F(1,23)=0.14,
p=0.7), interaction (F(1,23) = 0.7, p =0.4)).
Accidental vs. Intentional: In the RTPJ, accidental and intentional harms were more
discriminable in NT than ASD participants, reflected in a significant Group x Pattern
interaction (F(2,36)=4.9, p=.03), with no main effect of Pattern (F(1,36)=1.7, p=0.2) or
Group (F(1,36)=.71, p=0.4). The same interaction was observed in one-to-one matched
subsets of participants (F(1,28)=5.9, p=.02, see Appendix). There was no Group x
Pattern interaction in any other region (all F<0.9, all p>0.3).
Behavioral and Neural Correlation
The correlation of pattern discrimination in RTPJ with behavioral responses was
significantly larger in NT than in ASD participants (full sample, difference of correlations,
z=3.0, p=0.003; matched subsets z=2.6, p=0.009).
Decoding true vs. false beliefs?
One open question about the current results is whether the stimulus feature that we
successfully decoded is actually accidental vs. intentional harm, or another feature of
the beliefs in the scenarios which is confounded with this distinction. Two features are
confounded with accidental versus intentional: (i) false beliefs / ignorance versus true
beliefs / knowledge, and (ii) good / neutral versus bad intentions. In all of our stimuli,
accidental harms were actions based on a false belief or ignorance about the outcome
of the act and a neutral or good intention, while intentional harms were actions based on
a true belief or foreknowledge about the outcome of the act and a negative intention.
To test which of these features is represented by the patterns of neural response, we
analyzed additional data from Experiments 2 and 3. In these experiments, in addition to
accidental and intentional harms, participants read about failed attempts to harm (i.e.
neutral outcome, false belief, negative intention) and neutral actions (i.e. neutral
outcome, true belief, neutral intention). Like the comparison between accidental and
intentional harms, this contrast thus holds the outcome constant, and manipulates
whether the belief is true or false and whether the intention is good/neutral or bad.
We found that even when combining data from Experiments 2 and 3 (for maximum
power, n=30), the patterns in response to failed attempts and neutral actions were not
different in RTPJ (within=0.74(0.10), across=0.78(0.08), t(29)=-0.8, p=0.8). These
results suggest that the pattern distinction observed in RTPJ for accidental vs.
intentional harms, in the same participants, is not just due to the difference between true
and false beliefs, or whether the intention is good/neutral or bad.
62 of 179
Discussion
A central aim of this study was to ask whether the difference between accidental and
intentional harms could be decoded from the pattern of neural response within theory of
mind brain regions. Across three experiments, using different stimuli, paradigms, and
participants, we found converging results: in neurotypical (NT) adults, stories about
intentional versus accidental harms elicited spatially distinct patterns of response within
the right temporo-parietal junction (RTPJ). Moreover, this neural response mirrored
behavioral judgments: individuals who showed more distinct patterns in RTPJ also
made a larger distinction between intentional and accidental harms in their moral
judgments. Notably, in the fourth experiment, this pattern was absent in adults with
ASD: we found no neural difference between accidental and intentional harms in pattern
or mean signal, and behavioral judgments showed a marginal negative correlation with
neural pattern.
MVPA discriminates accidental and intentional harms in RTPJ of
neurotypical adults
The current results suggest that the RTPJ contains an explicit representation that
distinguishes intentional from accidental harms. This representation was apparent in
reliable but distinct spatial patterns of activity. The current results are among the first to
extend this method to high-level cognition and abstract stimulus features (see also
Fedorenko et al., 2012; Hampton & O'Doherty, 2007; Peelen et al., 2006; Tusche, Bode,
& Haynes, 2010). In particular, this is the first evidence of a feature of mental state
reasoning explicitly represented in a ToM brain region.
The convergence across experiments provides strong evidence that intentional and
accidental harms can be discriminated, using MVPA, in RTPJ. Designed to test a series
of separate questions, the three experiments differed in story content, voice of the
narrative (2nd or 3rd person), tense (past or present), the severity of harm caused (mild
in Exp.1, extreme in Exp.2, unspecified in Exp.3), the order and timing of the story
segments, the number of stories per condition, and the participants’ explicit task.
Perhaps most importantly, the cues to intent were different across experiments. In
Experiment 1, the same mental state content (e.g., your cousin’s allergy to peanuts)
was described as known or unknown (e.g., “you had no idea” vs. “you definitely knew”).
By contrast, in Exp.2 and 3, sentences with the same syntax and mental state verbs
were used to describe beliefs with different content (e.g. “Steve believes the ground
beef is safe / rotten”). Nevertheless, the spatial pattern of response was reliable and
distinct for intentional versus accidental harms specifically in the RTPJ, across all three
experiments. The converging results across experiments suggests that, rather than
being driven by superficial stimulus features or task demands, the distinct neural
patterns reflect an underlying distinction in the representation of accidental and
intentional harms.
The pattern difference found in the current work suggests that the distinction between
intentional and accidental harm is encoded in the neural representation in RTPJ.
Interestingly, the pattern in RTPJ did not distinguish between true and false beliefs, or
63 of 179
between negative and neutral intentions, in the context of non-harmful acts (i.e. the
difference between neutral acts and failed attempts to harm), suggesting that the
representations underlying the difference between accidental and intentional harm are
relatively specific.
This evidence that one feature of mental states is explicitly represented in the neural
pattern opens the door to many future studies. Many other features may also be
decodable. For example, patterns of response in theory of mind brain regions may
discriminate between attributing beliefs that are justified or unjustified (Young, Nichols, &
Saxe, 2010c), plausible or crazy (Young, Dodell-Feder, & Saxe, 2010b), attributed to
friends or enemies (Bruneau & Saxe, 2012), “constrained” or “open-ended” (Jenkins &
Mitchell, 2010), or first-order or higher-order (Koster-Hale & Saxe, 2011). Important
challenges for future research will include (i) determining the full set of mental state
features that can be decoded from the RTPJ response using MVPA, and (ii) identifying
the features of social stimuli that can be decoded from other regions within and beyond
the theory of mind network.
Pattern discrimination of intentional versus accidental harm is
correlated with moral judgment
The distinctness of the spatial patterns in the RTPJ was correlated with individuals’
moral judgments. Neurotypical adults differed in the amount of blame they assign to
accidental harms: some weighed intent more strongly (i.e., forgive more based on
innocent intentions), while others weighed outcome more strongly (i.e., blamed more
based on bad outcomes, Nichols & Ulatowski, 2007; Young & Saxe, 2009a). These
differences in moral judgment were predicted by individual differences in neural pattern
discriminability in the RTPJ, and more weakly in the LTPJ. Though Experiment 1 used a
blameworthiness scale (“How much blame should you get?”), and Experiment 2 used a
permissibility scale (“How permissible was Steve’s action?”), the same result emerged
in both studies: the individuals who encoded the difference between accidental and
intentional harms most strongly in their RTPJ also showed the greatest difference in
their moral judgments of these acts. These findings were corroborated by the wholebrain searchlight results: in the participants whose moral judgments were most sensitive
to the difference between intentional and accidental harm, only a region within RTPJ
discriminated between intentional and accidental harms.
Note that mental states (e.g., beliefs, intentions) represent just one of many inputs to
moral judgment. Decisions about moral blame and permissibility depend on many
features of the event, including the agent’s beliefs and desires (Cushman, 2008), the
severity of the harm (Greene, Sommerville, Nystrom, Darley, & Cohen, 2001), the
agent’s prior record (e.g., Woolfolk, Doris, & Darley, 2006), the means of the harm
(Cushman, Young, & Hauser, 2006; Greene, Nystrom, Engell, Darley, & Cohen, 2004),
the external constraints on the agent (e.g., coercion, self-defense, Woolfolk et al., 2006;
Young & Phillips, 2011), and more (e.g., Young et al., 2010c). Thus, activity in the RTPJ
reflects the representation of one input to moral judgment, rather than the judgment
itself.
64 of 179
In addition to finding distinct neural patterns in RTPJ, we also found a difference in the
magnitude of response when participants were reading about intentions and making
moral judgments: participants showed a higher level of RTPJ activity for accidental
harms relative to intentional harms. The current results converge with, and extend, prior
reports that RTPJ is recruited in the face of mitigating mental state information such as
innocent intent (Buckholtz et al., 2008; Yamada et al., 2012; e.g., Young & Saxe,
2009b). We find the effect in response magnitude is independent of the difference in
neural pattern, suggesting that these effects are sensitive to different aspects of RTPJ
function. A higher magnitude of response to accidental harms suggests that NT adults
use mental state reasoning in their moral judgments more when faced with relevant
information about the mind of the perpetrator. The stable difference in neural pattern
suggests that additionally the distinction between intentional and accidental harm is
encoded in the neural representation in RTPJ.
High functioning adults with ASD show atypical neural activity in
RTPJ
In ASD adults, we found distinct patterns for harmful and neutral acts in all theory of
mind regions, but no distinction between intentional and accidental harms in the RTPJ
or any other region. Similarly, mean signal did not differentiate between accidental and
intentional harms in any region. Finally, unlike in NT adults, behavioral responses
showed a marginal negative correlation with neural discrimination in the RTPJ,
The results of the current study match the behavioral profile of high-functioning
individuals with ASD. Individuals with ASD do not have impairments in moral judgment
as a whole: children with ASD make typical distinctions between moral and conventional
transgressions (Blair, 1996), and between good and bad actions (Leslie, Mallon, &
DiCorcia, 2006). However, they are delayed in using information about innocent
intentions to forgive accidents (Grant, 2005). Furthermore, even very high-functioning
adults with ASD, who pass traditional tests of understanding (false) beliefs, neglect
beliefs and intentions in their moral judgments compared to NT adults (Moran et al.,
2011; Yang et al., 2013); although in the current sample, this effect was observed more
strongly in z-scored behavioral data (Appendix and Baez et al., 2012).
Two prior papers have used MVPA to study neural mechanisms of social cognition
(though not specifically theory of mind) in ASD. Gilbert et al. (2008) measured the
magnitude and pattern of activity in MPFC while participants performed a task that did
or did not elicit thinking about another person. The magnitude of response during the
two tasks was equivalent in participants with ASD and controls, but participants with
ASD showed a less reliable and distinct pattern of activation in the MPFC during the
‘person’ task. Coutanche and colleagues (2011) measured the pattern and magnitude of
response in ventral visual areas while viewing faces versus houses. Again, the
magnitude of response in these regions was not different in typical adults and those with
ASD, but individuals with ASD showed less discriminable patterns in response to faces,
a social category, compared to houses. Disorganization of the pattern of activity for
faces versus houses was correlated with ASD symptom severity.
65 of 179
Broadly consistent with this prior work, the current results suggest that ASD affects the
organization (i.e., pattern) of information in theory of mind brain regions. One concern
might be that the reduced pattern information is merely a result of more noisy or
heterogeneous neural responses in ASD (e.g., Dinstein et al., 2010). Our results do not
favor this interpretation: rather than lower within-condition spatial correlations (i.e.,
noisier or more idiosyncratic responses), we found numerically higher within-condition
correlations in participants with ASD. That is, participants with ASD seemed to show a
reliable pattern of response in the RTPJ in response to all harmful acts, regardless of
whether the act was intentional or accidental.
We also found a difference between groups in the magnitude of response. NT adults
showed a higher response to accidental than intentional harms in the RTPJ, suggesting
increased activity the face of mitigating mental state information. In contrast, ASD adults
showed equal activation to both types of stories, suggesting that accidental harms did
not elicit more consideration of mental states than intentional harms. Given the
independence of the observed differences in magnitude and pattern, our results provide
converging evidence from two different types of analyses of atypical representations of
intentional versus accidental harms in the RTPJ of adults with ASD.
Prior studies investigating the magnitude of response in theory of mind regions of
individuals with ASD have found inconsistent results. Some studies suggest that theory
of mind regions are hypo-active (i.e., produce a smaller or less selective response,
(Kennedy & Courchesne, 2008; Lombardo et al., 2011), while other studies find no
difference between ASD and NT individuals (Gilbert et al., 2008; Nieminen-von Wendt et
al., 2003), and still others find the opposite pattern, hyper-activation, in ASD (Mason,
Williams, Kana, Minshew, & Just, 2008; Tesink et al., 2009; Wang, 2006). Studies using
tasks that elicit spontaneous or implicit social processing may be more likely to find
hypo-activation (Adolphs, 2001; Groen et al., 2010; Wang, 2006). Because increases in
magnitude may reflect either successful representation of a stimulus, or effortful but
ineffectual processing, differences in the magnitude of response between groups can be
difficult to interpret.
MVPA may therefore offer a sensitive tool for measuring neural differences in ASD that
are related to social impairments. Note, though, that due to the demands of the task and
scanning environment, the present ASD participants (as in previous task-oriented
neuroimaging studies e.g., Brieber et al., 2010; Koshino et al., 2007), and (see Philip et
al., 2010 for a review) are very high-functioning, which may limit the generalizability of
the results to lower-functioning individuals. Nevertheless, the individuals in the current
study do experience disproportionate difficulties with social interaction and
communication; therefore, the current results provide a window into the neural
mechanism underlying these difficulties.
Conclusion
In summary, using multivoxel pattern analyses across four experiments, we found that
(i) the difference between accidental and intentional harms is linearly decodable from
stable and distinct spatial patterns of neural activity in RTPJ; (ii) individual differences in
neural discrimination in RTPJ predict individual differences in moral judgment; and (iii)
66 of 179
these neural patterns are not detectable in high functioning adults with ASD.
Considerable neuroimaging work on theory of mind suggests that the RTPJ plays some
role in thinking about others’ thoughts; the current evidence suggests that one aspect of
this role is to make explicit, in the population response of its neurons, features of beliefs
that are most relevant for inference and decision-making — for example, encoding the
intent behind a harmful act.
Acknowledgments
This material is based upon work supported by the National Institute of Health under
grant 1R01MH096914-01A1, the Simons Foundation, the National Science Foundation
under grant 095518, a John Merck Scholars Grant, the Dana Foundation, and a
National Science Foundation Graduate Research Fellowship, grant 0645960. We thank
our participants, Lee Marvos and Caitlin Malloy for recruiting help, Nancy Kanwisher, Ev
Fedorenko, Hilary Richardson, Nick Dufour, members of the MIT Saxelab, the BC
Morality Lab, and the Brown Moral Psychology Research lab for useful comments and
discussion. Data were collected at the Athinoula A. Martinos Imaging Center at the
McGovern Institute for Brain Research, MIT.
Appendix
SI Behavioral Results: Results from Raw Behavioral Responses
In the main text, we report behavioral data and statistics that have been z-scored to
account for variability in the participants’ use of the rating scales. Here we provide the
same analyses, using the untransformed button press responses. We find largely
converging results using untransformed behavioral data.
Experiment 1: From no blame at all (1) to very much blame (4), participants judged
intentional harms (3.4 ± 0.09) to be more blameworthy than accidental harms (1.5 ±
0.08; t(19)=19, p<0.0001), both of which were judged to be worse than neutral stories
(1.1 ± .03, t(19)=30, p<0.0001; t(19)=6.9, p<0.0001).
Experiment 2: Replicating the results in Exp.1, on a scale of Forbidden (1) to
Permissible (3), participants judged intentional harms (1.1 ± .04) as less permissible
than accidental harms (2.2 ± .11; t(9)=11, p<0.001).
Experiment 3: Participants in Exp.3 did not make moral judgments of the scenarios, but
instead answered true/false questions about the content of the fact.
Experiment 4: When making moral judgments, ASD participants, like NT participants
from Exp.1, judged intentional harms (3.4 ± 0.18) to be worse than accidental harms
(1.9±0.17; t(10)=7.5; p<0.0001), both of which were judged to be worse than neutral
stories (1.2±.06; t(10)=10.7, p=8.3e-07; t(10)=4.3, p=0.0016).
Group comparison: A mixed effects ANOVA crossing Group (NT in Exp. 1, ASD in Exp.
4) by Condition (accidental, intentional raw behavioral ratings) yielded a main effect of
67 of 179
Condition (F(1,29)=381.1, p<0.0001), with no effect of Group F(1,29)=2.7, p=0.1, and
no interaction (F(1,29)=1.1, p=0.3). The planned comparison t-test revealed that ASD
adults assigned more blame for accidental harms than NT adults (t(14.2)=1.7, p=0.05,
one-tailed; see Moran et al., 2011).
Behavioral and Neural Correlation: In Exp.1 and 2, NT participants provided moral
judgments of each scenario in the scanner, allowing us to determine whether behavioral
responses were related to the spatial pattern of the neural response in RTPJ or any
other region. For each participant, we calculated the difference in moral judgments for
intentional versus accidental harms. We tested whether this difference score was
correlated, across participants, with the index of classification in each region (intentional
vs. accidental, within-condition correlation minus across-condition correlation). In both
experiments, we found that only in the RTPJ was the difference between intentional and
accidental harms in individuals’ moral judgments correlated with the neural classification
index (Exp. 1: r(17)=0.55; p=0.016; Exp. 2: r(8)=0.71; p=0.02). The correlation between
neural pattern and behavior remained significant after combining the data from both
experiments, using the moral judgments converted to the same scale 7 (r(27)=0.56,
p=0.002). No other regions showed a significant correlation with behavioral judgments,
through there was a strong trend in LTPJ (r(27)=0.35, p=0.06).
Supplementary Results: Pairwise matched subsets of ASD and NT
In the main manuscript, we mostly focus on the comparison of the full sample of NT
(n=23) and ASD (n=16) participants. Here we report the results of the key analyses
when conducted in a subset of each group that were matched in pairs. These analyses
find the same key pattern of results; importantly, the group by condition interaction in the
neutral pattern of RTPJ remains significant.
Participants: Of the 16 ASD participants, we were able to create one-to-one matches
on gender, age, and IQ for 15 participants, resulting in two groups (NT and ASD), both
n=15, 13 male and 2 female. These pairs are matched in age (NT (mean ± SD)=30
years ±10; ASD=30 years ± 8; maximum difference between pairs 7 years, average
absolute difference 3.4 years, t(26.5)=0.26, p=0.8) and IQ (NT mean=121 ± 11.8;
ASD=120 ± 12.7, maximum difference 12 points, average absolute difference 4.2
points, t(27.1)=0.32, p=0.7). As in the full sample, ASD participants scored significantly
higher than NT participants on the AQ (NT (mean ± SD)=18.5 years ± 6.6; ASD=32.5 ±
7.2; t(20.2)=4.8, p < 0.0001).
Behavioral Results: Behavioral data were available for 11 ASD and 13 NT participants
from the matched subsets (NT: accidental -0.39 ± 0.05, intentional 1.17 ± 0.03; ASD:
accidental -0.23 ± 0.07, intentional 1.01 ± 0.09). In these subjects, a mixed effects
ANOVA crossing Group (NT, ASD) by Condition (accidental, intentional z-scored ratings)
yielded a main effect of Condition (F(1,22)=304, p<0.0001), and a marginal Group by
7
Raw Data: Behavioral data (key presses) from Experiments 1 and 2 were translated to the same scale
for comparison. Specifically, responses from Experiment 2 were inverted to make “forbidden / very much
blame” the top end of the scale for all experiments by subtracting each response from 4 (e.g., a rating of 1
becomes a rating of 3), and then scaling linearly from a 3-point scale to a 4-point scale.
68 of 179
Condition interaction (F(1,22)=3.93, p=0.059). A planned comparison (see Moran et al.,
2011) t-test revealed that, in the matched subsets, ASD adults showed the same trend
of assigning more blame for accidental harms than NT adults (t(22)=1.6, p=0.06, onetailed).
Motion and Artifact: The subsets of NT and ASD participants did not differ in the
number of motion artifacts per run (NT: 4.3 ± 1.2; ASD: 3.6 ± 1.3, t(27.9)=0.35, p=0.7),
in the number of global signal outliers per run (NT: 2.5 ± 0.4; ASD: 2.7 ± 0.3,
t(26.1)=0.50, p=0.6), or in the total vector translation (NT: 0.25 ± 0.02), ASD: 0.23 ±
0.02, t(28)=0.73, p=0.47.
Voxel-Wise Pattern Results: Harm vs. Neutral: As in the full sample, the neural pattern
in the RTPJ of both NT and ASD participants discriminated harms from neutral actions
(NT: within=1.3±0.11, across=1.2±0.13, t(14)=2.0, p=0.03; ASD: within=1.5±0.14,
across=1.2±0.16, t(14)=3.7, p=0.001). A Group (ASD, NT) x Pattern (within, across)
ANOVA revealed that matched NT and ASD participants show equally robust neural
discrimination in response to harmful versus neutral actions in their RTPJ, with a main
effect of Pattern (F(1,28)=16.7, p=0.0003), no effect of Group (F(1,28)=0.3, p=0.5), and
no interaction (F(1,28)=1.7, p=0.2).
Accidental vs. Intentional: As in the full sample, the neural pattern in the RTPJ of NT
participants discriminated accidental from intentional harms (within=1.1±0.14,
across=0.99±0.15, t(14)=2.2, p=0.02), while matched ASD participants did not
(within=1.3±0.11, across=1.3±0.13, t(14)=1.1, p=0.8). Matched NT participants
discriminated between accidental and intentional harms to a greater extent than ASD
participants, reflected in a significant Group x Pattern interaction (F(1,28)=5.9, p=.02),
with no main effect of Group (F(1,28)=1.93, p=0.18) or Pattern (F(1,28)=1.53, p=0.23).
Behavioral and Neural Correlation: The correlation of pattern discrimination in RTPJ
with behavioral responses was significantly larger in NT (r(11)=0.56; p=0.047) than in
ASD participants (r(9)=0.54, p=0.08; z=2.6, p=0.009). Note that the marginal effect in
ASD is in the reverse direction: stronger neural discrimination in individuals with more
similar behavioral moral judgments.
Independent effects of magnitude and pattern
In neurotypical adults, we observed differences between accidental and intentional
harms in the pattern of response over the whole trial and in the magnitude response in
the final phase. Specifically in the RTPJ, both of these effects contrasted with the
pattern observed in ASD adults, with significant Condition by Group interactions. When
interpreting these results it is important to know whether these two different ways of
measuring the same effect or converging evidence of two different aspects of atypical
RTPJ function in ASD. To address this question, we asked whether these two effects
were correlated, across individuals. We found these two effects were not correlated
across either NT (r(20)=0.11, p=0.62) or ASD (r(14)=-0.16, p=0.55, suggesting that the
effects are sensitive to different aspects of RTPJ function. Moreover, our analysis
techniques are sensitive to different features of the data: because correlations are not
sensitive to overall magnitude, the pattern analyses used here are not driven simply by
differences in the magnitude (see discussion in the next section).
69 of 179
A note on separating the phases of Experiment 1
Over the full duration of the trial, in the RTPJ we found no difference in the magnitude of
response to intentional versus accidental harms (accidental: 0.26±0.05; intentional:
0.22±0.05; accidental vs. intentional: t(21)=1.4, p=0.18). By contrast, we could reliably
decode the difference between these two conditions from the spatial patterns of the
betas (within=1.2±0.12, across=1.1±0.12, t(21)=2.2, p=0.02). When we tried to
separately estimate the response to only the second phase of the trial (Intent, Decision),
we found a higher magnitude of response in RTPJ for accidental than intentional harms
(accidental: 0.27±0.06; intentional: 0.22±0.07; accidental vs. intentional: t(21)=2.95,
p=0.008). However, we could not reliably the decode the difference between these two
conditions from the spatial patterns of the betas (within=0.98±0.12, across=0.92±0.11,
t(21)=1.3, p=0.11).
These results illustrate, first, that the current techniques for estimating the magnitude
and pattern of neural responses are independent. Because correlations depend only on
the spatial pattern of the response (i.e. which voxels show relatively higher betas, and
which show relatively lower betas), it is possible to find a pattern difference in the
absence of a magnitude difference, and it is also possible to find a magnitude difference
in the absence of a pattern difference. This is a difference between the current MVPA
technique (focusing exclusively on spatial patterns based on Haxby, 2001) and other
machine learning algorithms used commonly in the literature (e.g support vector
machines).
Second, these results illustrate the different sensitivity of analyses based on average
percent signal change, versus estimated betas of modeled hemodynamic response
functions. Because there was no jitter in the timing (i.e. the onset of the Intent segment
was always perfectly predictable from the onset of the story) separate regressors for the
first and second phases of each trial are partially co-linear. As a result, the withincondition correlations of betas (a measure of our ability to estimate reliable betas) were
notably lower for the same subjects in the last-segment-only (e.g. RTPJ within=0.98
±0.12), compared to the full trial (e.g. RTPJ within=1.2±0.12). Note that these
correlations reflect the reliability of the spatial pattern across trials, not the magnitude of
the betas. Because the discriminating information all occurs in the second phase of the
trial, successful decoding from the whole trial but not the second phase alone must
reflect the unreliable estimates of second-phase betas.
70 of 179
Chapter 3.
Thinking about Seeing: perceptual sources of
knowledge are encoded in the theory of mind brain
regions of sighted and blind adults 8!
Blind people’s inferences about how other people see provide a window into
fundamental questions about the human capacity to think about one another’s
thoughts. By working with blind individuals, we can ask both what kinds of
representations people form about others’ minds, and how much these
representations depend on the observer having had similar mental states
themselves. Thinking about others’ mental states depends on a specific group of
brain regions, including the right temporo-parietal junction (RTPJ). We investigated
the representations of others’ mental states in these brain regions, using multivoxel
pattern analyses (MVPA). We found that, first, in the RTPJ of sighted adults, the
pattern of neural response distinguished the source of the mental state (did the
protagonist see or hear something?) but not the valence (did the protagonist feel
good or bad?). Second, these neural representations were preserved in congenitally
blind adults. These results suggest that the temporo-parietal junction contains
explicit, abstract representations of features of others’ mental states, including the
perceptual source. The persistence of these representations in congenitally blind
adults, who have no first-person experience with sight, provides evidence that these
representations emerge even in the absence of relevant first-person perceptual
experiences.
8
A version of this chapter was published as Koster-Hale, J., Bedny, M., & Saxe, R. R. (2014). Thinking
about seeing: Perceptual sources of knowledge are encoded in the theory of mind brain regions of
sighted and blind adults, Cognition 133(1), 65–78.
71 of 179
Introduction
Imagine a friend tells you that last night, looking out the window onto a dark, rainy
street, she saw her boyfriend get into a car with a strange woman, and drive away. Your
reaction will depend on many inferences about her thoughts and feelings. You will
recognize that she believes her boyfriend is being unfaithful, and feels betrayed. You
might also note the source of her belief, and question how clearly she could see at a
distance and in the dark. Perhaps she was mistaken about the identity of the man
getting into the car, or about the driver; maybe the driver was actually her boyfriend’s
sister. Knowing how she got her information might strongly affect how you reason about
her beliefs and experiences — do you yourself believe that her boyfriend is being
unfaithful? How strongly do you think she believes it? What is she likely to do next?
Now imagine that you are congenitally blind. Would your inferences be any different?
Clearly, a blind adult would understand the emotional toll of discovering a lover’s
possible betrayal, but could a blind person make the same inferences about the visual
source of the discovery? How much would a blind person understand about the
experience of seeing a familiar person and a strange woman, from afar, in the dark?
Blind people’s inferences about how other people see provide a window into a
fundamental question about the human capacity to think about one another’s thoughts:
what are the mechanisms used to think about someone else’s mind? One possibility is
that we think about someone’s experience by invoking our own relevant experiences
and sensations. In this view, thinking about someone else’s experience of seeing
requires (among other things) a first-person experience of sight. In contrast, if people
use an intuitive “theory” of mind, composed of relationships among abstract mental
state concepts, to reason about others’ experiences, then experience of sight is not
always necessary for reasoning about seeing. In many cases, these views are hard to
disentangle; however, they predict different outcomes in blindness. If first-person
experience is necessary to understand others’ experiences, blind people should have
only a fragmentary, limited, or metaphorical understanding of seeing. By asking how
blind individuals represent other’s experiences of sight, we can place important limits on
our theories of mental state inference: to what extent does theory of mind depend on
the observer having had similar sensations, experiences, beliefs, and feelings as the
target?
The first possibility is that people understand another’s mind by trying to replicate it, by
imagining themselves in a similar situation, or by re-experiencing a similar past event of
their own lives. You would understand your friend’s feelings of sadness by imagining
your own feelings in response to a lover’s betrayal, recreating in your emotional system
a version of your friend’s experience. Similarly, you would understand your friend’s
experience of seeing her boyfriend get into the car by recreating the visual scene in
your own mind’s eye (Gordon, 1989; Stich & Nichols, 1992). Understanding others’
minds thus depends on the observer having experienced a relevantly similar mental
state to the target (Gallese & Goldman, 1998; Goldman, 1989; 2006; Gordon, 1989;
Harris, 1992; Nichols, Stich, Leslie, & Klein, 1996; Stich & Nichols, 1992). Simulationbased accounts do not necessarily posit that people can only think about exactly those
72 of 179
experiences that they themselves have had; rather, mental state representation could
be a composition of one’s existing relevantly similar first-person experiences, composed
flexibly to simulate a novel experience. Still, because simulation depends on similar
experiences, the extent to which we can simulate the minds of others depends on “the
interpersonal sharing of the same kind of neural and cognitive resources. When this
sharing is limited (or even missing), people are not fully able (or are not able at all) to
map the mental states or processes of others because they do not have suitable mental
states or processes to reuse” (Gallese & Sinigaglia, 2011). Because congenitally blind
people lack the mental states and processes involved in seeing, their representations of
sight are predicted to be limited or unreliable.
Neuroimaging experiments provide evidence that first-person sensorimotor
representations are “reused” during observation of others’ actions and sensations.
Similar brain regions are recruited when experiencing physical pain compared to
observing another person experience similar pain (e.g., Botvinick et al., 2005;
Immordino-Yang, McColl, Damasio, & Damasio, 2009; Singer et al., 2004); and when
experiencing a tactile sensation compared to observing another person being touched
in the same way (Blakemore, 2005; Keysers:2004dj Gazzola & Keysers, 2009). More
importantly, neural activity during some observation tasks depends on the observer’s
own specific first-person experiences. For example, motor activation in dancers during
observation of dance moves is enhanced for specific movements that the observers
themselves have frequently executed (Calvo-Merino, Glaser, Grèzes, Passingham, &
Haggard, 2005; Calvo-Merino, Grèzes, Glaser, Passingham, & Haggard, 2006). If this
type of reuse extends to mental state representation, typical representations of other’s
experiences of seeing should depend on, or be profoundly affected by, having first seen
yourself.
However, many authors (Gopnik & Meltzoff, 1997; Gopnik & Wellman, 1992; Perner,
1993; Saxe, 2005) have suggested an alternative mechanism for understanding other
minds: namely, that people have an intuitive theory of other minds. An intuitive theory
includes causal relations among abstract concepts (like beliefs and desires), and can be
learned from many sources of evidence, not limited to first-person experiences. One
source of evidence children could use to build a theory of mind is the testimony of
others: verbal labels and descriptions of mental states and experiences, often including
mental state verbs like think and see (Harris, 2002b; 1992). Thus a congenitally blind
child, growing up in a world full of sighted people, might develop an intuitive theory that
includes concepts of vision, to explain everyone else’s behavior (e.g. reacting to objects
at a distance) and testimony (e.g. saying “I see your toy on the top shelf!”). This intuitive
theory would then allow a blind child to predict how a sighted person would act in a
given environment, and what that person would be likely to infer based on what she
could see.
To test these theories, we investigated how blind people think about sight. Observation
and behavioral studies suggest that even young blind children know that other people
can see with their eyes, and can understand basic principles of vision: e.g. that objects
can be seen from a distance and are invisible in the dark (Bigelow, 1992; Landau &
Gleitman, 1985; Peterson, & Webb, 2000). By adulthood, congenitally blind people
know the meanings of verbs of sight, including fine-grained distinctions, such as the
73 of 179
difference between verbs like peer, gaze, and gawk (Landau & Gleitman, 1985; Lenci,
Baroni, Cazzolli, & Marotta, 2013, Koster-Hale et al., in prep). Blind adults are thus
sensitive to subtle distinctions in how sighted people gather information visually.
This behavioral evidence alone, however, cannot answer the question of whether a
blind person uses the same cognitive mechanisms as a sighted person to understand
sight. Any surface similarity between a blind and sighted person’s verbal descriptions of
sight could be the product of compensatory mechanisms in the blind. For example,
some authors have suggested that blind people mimic the words used by sighted
people, without being able to fully access their meaning (so-called “verbalisms,” Rosel,
Caballer, Jara, & Oliver, 2005) or integrate them into their conceptual understanding.
Thus, a blind person who hears the sentence “I saw my boyfriend getting into the car
from the window,” may have only a limited or metaphorical understanding of the
experience it describes.
This methodological challenge affords an opportunity for cognitive neuroscience:
functional neuroimaging can provide an online, unobtrusive measure of ongoing
psychological processes and thus offers an alternative strategy to ask if two groups of
people are performing a cognitive task using similar or different mechanisms. Here we
use neuroimaging to ask whether blind and sighted people rely on similar cognitive
mechanisms when they reason about seeing.
Previous studies have shown that thinking about someone else's thoughts (including
those based on visual experiences, Bedny et al., 2009) increases metabolic activity in a
specific group of brain regions often called the 'mentalizing' or theory of mind network.
These regions include the medial prefrontal cortex (MPFC), the precuneus (PC), and
the bilateral temporal-parietal junction (TPJ), (e.g., Aichhorn et al., 2009; Bedny,
Pascual-Leone, & Saxe, 2009; Lombardo, Chakrabarti, Bullmore, Baron-Cohen, &
Consortium, 2011; Mason & Just, 2011; Rabin, Gilboa, Stuss, Mar, & Rosenbaum, 2010;
Saxe & Kanwisher, 2003; Saxe & Powell, 2006; Spunt, Satpute, & Lieberman, 2010).
However, very little is known about which aspects of mental states are represented in
these brain regions. Mental experiences have many features, including the content
(“boyfriend in stranger’s car”), the valence (very sad), and the modality or source of the
experience (seen from afar). Neurons, or groups of neurons, within theory of mind brain
regions may represent any or all of these features. Thus, two important questions
remain open. First, in sighted people, do neurons in any theory of mind region
specifically represent the perceptual source of another person’s belief? Second, if so,
are similar representations present in blind people?
A powerful approach for understanding neural representation is to ask which features of
a stimulus are represented by distinct subpopulations of neurons within a region. For
example, within middle temporal visual, subpopulations of neurons that respond to
visual stimuli moving orientations are spatially organized at a fine spatial scale. Although
no individual fMRI voxel (which includes hundred of thousands of neurons) shows an
orientation selective response, and therefore overall early visual cortex shows equal
average magnitude of response to all orientations, it is possible to detect reliably distinct
spatial patterns of response across cortex that do distinguish between orientations
(Kamitani & Tong, 2006). This technique of looking for reliable spatial patterns of fMRI
74 of 179
response within a brain region is called multivoxel pattern analysis (MVPA; (Haynes &
Rees, 2006; Kriegeskorte & Bandettini, 2007; Norman, Polyn, Detre, & Haxby, 2006).
MVPA can reveal how stimulus categories are processed within a functional region
(Peelen et al, 2006; Haynes & Rees, 2006). MVPA has been successfully used to probe
the neural basis of many different types of representation, including subjectively
perceived directions of motion in ambiguous stimuli, semantic category, emotional
affect, and intent when causing harm (Chapter 2, Koster-Hale, Saxe, Dungan, & Young,
2013; Mahon & Caramazza, 2010; e.g., Norman et al., 2006; Peelen, Wiggett, &
Downing, 2006; Serences & Boynton, 2007).
Here, we ask if source modality (seeing vs hearing) is a relevant feature of the neural
representations of mental states. If one set of neurons responds more to stories about
seeing than hearing, while another (partially distinct) set responds more to stories about
hearing than seeing, we have evidence that seeing and hearing are being represented
in different ways in that brain region. Measuring the average magnitude of activity in
theory of mind brain regions cannot be used to address this question, because the
average magnitude may obscure distinct subpopulations of neurons within the regions.
We therefore used multivoxel pattern analysis (MVPA) to look for differences in the
neural response to stories about hearing and seeing.
In this study, we first asked whether stories about someone else’s hearing and seeing
experiences evoke different spatial patterns of response within theory of mind brain
regions in sighted individuals. Because very little is known about how theory of mind is
represented, cognitively or neurally, this is itself a fundamental question about mental
state representation. We then asked whether the same patterns are observed in
congenitally blind people. Finding these patterns would provide support for the notion
that blind individuals represent mental experiences of seeing in a qualitatively similar
manner to sighted individuals. For comparison, we also tested whether theory of mind
regions encode a feature of mental states that should not differ between sighted and
blind people: the valence (feeling good versus bad).
75 of 179
Methods
Participants
Thirteen sighted members of the larger MIT community participated (8 women; mean
age ± SD, 52 years ± 16), all with normal or corrected-to-normal vision. Ten blind
individuals participated (5 women; mean age ± SD, 50 years ± 7). Nine blind
participants were born blind and one lost sight between the ages of 2 and 3 years. Due
to possible effects of early vision, this person was excluded from all analyses. All blind
participants reported having at most faint light perception and no pattern discrimination.
None were able to perceive shapes, colors, or motion.
One blind participant was ambidextrous and one was left-handed; one sighted
participant was left-handed. All were native English speakers and gave written informed
consent in accordance with the requirements of the Institutional Review Board at MIT.
Participants were compensated $30/hour for their time.
Comparison to Bedny et al. (2009)
These data have previously been published, analyzing the magnitude but not the
pattern of response in each region, in Bedny et al. (2009). Bedny et al. (2009) found that
theory of mind regions showed equally high responses to stories about hearing and
seeing, in both sighted and blind participants. However, measuring the average
magnitude across the entire region may obscure distinct subpopulations of neurons
within a region. Specifically, the equally high magnitude of response to stories about
seeing and hearing observed by Bedny et al. (2009) is consistent with three
possibilities: (i) neurons in theory of mind brain regions do not distinguish between
hearing and seeing in blind or sighted people, (ii) distinct subpopulations of neurons
within theory of mind brain regions respond to stories about seeing versus hearing in
both sighted and blind participants, reflecting a common representation of the
perceptual source of other’s mental states, or (iii) distinct subpopulations of neurons
respond to seeing versus hearing in sighted but not blind individuals, reflecting different
representations in the two groups, depending on their first person experiences. The
current analyses allowed us to distinguish between these possibilities.
Note that the current results exclude one blind participant who was included in the
previous paper: this participant was born with cataracts and had some light and shape
perception during the first ten years of his life. To be conservative, we exclude his data
from the current analyses.
Finally, the previous paper included an analysis of reaction time to the behavioral
portion of the task (judging how good or bad the protagonist felt at the end of the story),
looking at the seeing events, hearing events, and additional control events, all collapsed
across valence. In this paper, we are also treating valence as a dimension of interest in
the neural data, and so break down the behavioral data by both modality and valence to
report reaction time and rating data.
76 of 179
fMRI Protocol and Task
Source task: Participants heard 32 stories (each 13 sec long): four stories in each of
eight conditions. Stories in the four conditions of interest described a protagonist’s
mental experiences, characterized by both a modality-specific source (something seen
vs heard) and a modality-independent valence (whether the protagonist felt good or
bad). The “seeing” stories described the protagonist coming to believe something as a
result of a visual experience, such as seeing a friend’s worried face or recognizing
someone’s handwriting. The “hearing” stories described the protagonist coming to
believe something as a result of an auditory experience, such as hearing a friend’s
worried voice or recognizing someone’s footsteps. The stories with negative valence
described an event that would make the protagonist feel bad, such as receiving criticism
or losing a game; the stories with positive valence described a good event, such as
receiving praise or winning a game (Figure 3.1). The remaining four control conditions,
which did not describe mental experiences, are not analyzed here (see Bedny et al.,
2009 for details).
Figure 3.1: Sample stimuli.
77 of 179
For each narrative context (e.g. a job interview, dinner with parents-in-law, cleaning a
dorm room), we constructed four endings, one in each condition. Thus the stories
formed a matched and counterbalanced 2x2 (seeing vs hearing, positive vs negative)
design. Individual participants saw each context only once; every context occurred in all
conditions, across participants. Word count was matched across conditions (mean
length ± SD, 32 words ± 4). Stories were presented in a pseudorandom order, condition
order was counterbalanced across runs and subjects, and no condition was immediately
repeated. Rest blocks of 11 sec were presented after each story. Eight stories were
presented in each 4 min 34 sec run. The total experiment, four runs, lasted 18 min 18
sec.
After each story, participants indicated whether the main character in the story felt very
bad, a little bad, a little good, or very good, using a button press (1-4). Reaction time
was measured from the onset of each question.
Theory of Mind Localizer task: Participants also completed a theory of mind and
language localizer task. Participants listened to 48 short verbal stories from two
conditions: 24 stories requiring inferences about mental state representations (e.g.,
thoughts, beliefs) and 24 stories requiring inferences about physical representations
(e.g., maps, signs, photographs). These conditions were similar in their metarepresentational and logical complexity but differ in whether the reader is building a
representation of someone else’s mental state (See Saxe & Kanwisher, 2003, DodellFeder et al. 2011, and Bedny et al., 2009 for further discussion). After each story,
participants answered a true/false question about the story. As a control condition,
participants listed to 24 blocks of “noise,” unintelligible backwards speech created by
playing the stories backwards. The task was performed in 6 runs with 12 items per run
(4 belief, 4 physical, and 4 backward-speech). Each run was 6 min and 12 sec long.
The stimuli for the localizer and both experiments were digitally recorded by a female
speaker at a sampling rate of 44,100 to produce 32-bit digital sound files.
Acquisition and Preprocessing
fMRI data were collected in a 3T Siemens scanner at the Athinoula A. Martinos Imaging
Center at the McGovern Institute for Brain Research at MIT, using a 12-channel head
coil. Using standard echoplanar imaging procedures, we acquired T1-weighted
structural images in 128 axial slices with 1.33 mm isotropic voxels (TR = 2 ms, TE =
3.39 ms), and functional blood oxygen level dependent (BOLD) data in 30 near-axial
slices using 3x3x4 mm voxels (TR=2 s, TE=40 ms, flip angle=90°). To allow for steady
state magnetization, the first 4 seconds of each run were excluded.
Data processing and analysis were performed using SPM2 (http://www.fil.ion.ucl.ac.uk/
spm) and custom software. The data were motion corrected, realigned, normalized onto
a common brain space (Montreal Neurological Institute (MNI) template), spatially
smoothed using a Gaussian filter (full-width half-maximum 5 mm kernel) and subjected
to a high-pass filter (128 Hz).
78 of 179
Motion and Artifact Analysis
To estimate motion and data quality, we used three measures for each participant:
mean translation was defined as the average absolute translation for each TR across
the x, y, and z plane; mean rotation was defined as the average absolute rotation per
TR across yaw, pitch, and roll; and number of outliers per run was defined as the
average number of time points per run in which (a) TR-to-TR head movement exceeded
2mm of translation, or (b) global mean signal deviated by more than three standard
deviations of the mean. One blind participant moved excessively during the scan (mean
translation of 1.2 mm, mean rotation of 1.4°, and mean outliers per run = 20.7,
compared to other blind participants’ means of 0.33 mm of rotation, 0.35° of rotation,
and 2.5 outliers), so his results were dropped from the analyses.
fMRI Analysis
All fMRI data were modeled using a boxcar regressor, convolved with a standard
hemodynamic response function (HRF). The general linear model was used to analyze
the BOLD data from each subject, as a function of condition. The model included
nuisance covariates for run effects, global mean signal, and an intercept term. A slow
event-related design was used. An event was defined as a single story, the event onset
was defined by the onset of the story sound file, and offset as the end of the story.
Functional Localizer: Individual subject ROIs
In each participant, functional regions of interest (ROIs) were defined in right and left
temporo-parietal junction (RTPJ, LTPJ), medial precuneus (PC), dorsal medial
prefrontal cortex (DMPFC 9), and ventral medial prefrontal cortex (VMPFC3). Each
subject’s contrast image (Belief > Photo) was masked with each of the regions’ likely
locations, using probabilistic maps created from a separate dataset.The peak voxel that
occurred in a cluster of 10 or more voxels significant at p<0.001 was selected. All voxels
within a 9mm radius of the peak voxel, individually significant at p<0.001, were defined
as the ROI.
Within-ROI Pattern Correlation Analysis
Split-half correlation-based MVPA asks whether the neural pattern in a region is
sensitive to a category-level distinction. Specifically, we ask whether the neural patterns
generated by items within a condition are more similar to each other (“within-condition
correlation”) than to the neural patterns generated by items in the other conditions
(“across-condition correlation”). If we find that within-condition correlations are reliably
higher than across-condition correlations, we can conclude that there are reliable neural
pattern generated by different items within a condition, and that these neural patterns
are distinct from one condition to another. Together, this suggests that the region is
sensitive to the category distinction — items within a category are coded in a similar
way, with distinguishable codes for different categories.
9
Note that in Bedny et al., 2009, DMPFC was called SOFC, and VMPFC was called OFC.
79 of 179
Here, we conducted within-ROI pattern analyses, independently testing for information
about belief source and valence in the regions identified in the independent functional
localizer. To compare seeing and hearing beliefs, we collapsed across good and bad
valence; to compare good
and bad valence, we
collapsed across seeing
and hearing.
Following Haxby et. al.
(2001), each participant's
data were divided into
even and odd runs
(‘partitions’) and then the
mean response (beta
value) of every voxel in
the ROI was calculated
for each condition.
Because each participant
read 8 stories about
hearing, seeing, feeling
good, and feeling bad,
each partition contained
the average response to 4
individual stories. The
"pattern" of response was
Figure 3.2: Split half MVPA analysis. The data from each condition the vector of beta values
(e.g. hearing and seeing) were divided by run, and for each across voxels within the
participant, we asked whether the within condition correlations (blue participants individual
arrows) were higher than the across conditions correlations (grey ROI. To determine the
arrows). Data from one blind participant: within-condition correlation w i t h i n - c o n d i t i o n
= 1.3; across-condition correlation = 0.8.
correlation, the pattern in
one (e.g. even) partition
was correlated with the pattern for the same condition in the opposite (e.g. odd)
partition; to determine the across-condition correlations the pattern was compared to the
opposite condition, across partitions (Figure 3.2).
For each condition pair (e.g. seeing vs hearing) in each individual, an index of
classification was calculated as the within-condition correlation (e.g.the correlation of
one half of the seeing stories to the other half of seeing stories, averaged with the
correlation of one half of the hearing to the other half of hearing stories) minus the
across-condition correlation (e.g. the correlation of seeing stories compared to hearing
stories). To allow for direct comparison of correlation coefficients, we transformed all r
values using Fisher’s Z transform. A region successfully classified a category of stimuli
if, across individuals, the within-condition correlation was higher than the acrosscondition correlation, using a Student's T complementary cumulative distribution
function.
80 of 179
This procedure implements a simple linear decoder. Linear decoding, while in principle
less flexible and less powerful than non-linear decoding, is preferable both theoretically
and empirically. A non-linear classifier can decode nearly any arbitrary feature contained
implicitly within an ROI, reflecting properties of the pattern analysis algorithm rather than
the brain, which makes successful classification largely uninformative (Cox & Savoy,
2003; DiCarlo & Cox, 2007; Goris & Op de Beeck, 2009; Kamitani & Tong, 2005;
Norman et al., 2006). Moreover, linear codes have been argued to be a more neurally
plausible way of making information available to the next layer of neurons (Bialek,
Rieke, Van Steveninck, & Warland, 1991; Butts et al., 2007; DiCarlo & Cox, 2007;
Naselaris, Kay, Nishimoto, & Gallant, 2011; Rolls & Treves, 2011).
Whole Brain Pattern Correlation Analysis (Searchlight)
To ask whether any other part of the brain encoded the difference between other
people's experiences of seeing and hearing, or between good and bad events, we used
a searchlight analysis to look across the brain. In the searchlight analysis, rather than
using a predefined ROI, a Gaussian kernel (14mm FWHM, corresponding
approximately to the observed size of the functional ROIs) was moved iteratively across
the brain. Using the same logic as the ROI-based MVPA, we computed the spatial
correlation, in each kernel, of the neural response (i.e. betas) within conditions and
across conditions. We then transformed the correlations using Fisher’s Z, and
subtracted the across-condition correlation from the within-condition correlation to
create an index of classification. Thus, for each voxel, we obtained an index of how well
the spatial pattern of response in the local region (i.e. the area centered on that voxel)
can distinguish between the two conditions. The use of a Gaussian kernel smoothly deemphasizes the influence of voxels at increasing distances from the reference voxel
(Fedorenko, Nieto-Castañón, & Kanwisher, 2012). We created whole brain maps of the
index of classification for each subject. These individual-subject correlation maps were
then subjected to a second-level analysis using a one-sample t-test (thresholded at
p<0.001, voxelwise, uncorrected).
81 of 179
Results
fMRI task: Behavioral Results
Sighted participants
We performed a 2x2 ANOVA, crossing valence (good/bad) and modality (seeing/
hearing), on both goodness ratings and reaction time data. Using a 1 (very bad) to 4
(very good) scale, sighted participants rated protagonists who experienced positive
events as feeling better than the protagonists experiencing negative events, with no
effect of modality and no interaction (hearing-bad: 1.84±0.14, hearing-good: 3.43±0.14,
seeing-bad: 1.88±0.12, seeing-good: 3.65±0.11; main effect of valence: F(1,12)=223,
p<0.001, partial η2=0.79; modality: F(1,12)=1.2 p=0.3, partial η2=0.02; modality by
valence interaction: F(1,12)=0.4, p=0.5, partial η2=0.01).
Sighted participants showed small, but significant effects of condition on reaction time
(measured from onset of the question), with a marginal main effect of modality and a
small but significant interaction (hearing-bad: 5.63±0.23, hearing-good: 5.2±0.21,
seeing-bad: 5.03±0.3, seeing-good: 5.28±0.23; main effect of modality: F(1,12)=4.4,
p=0.06, partial η2=0.02; valence: F(1,12)=0.3, p=0.5, partial η2<0.01; modality by
valence interaction: F(1,12)=6.8, p=0.002, partial η2=0.04). Post-hoc t-tests reveal that
sighted participants responded more slowly to stories about negative experiences
based on hearing than based on seeing (t(12)=3.3, p=0.007); there was no effect of
modality on responses to positive events (t(12)=0.47, p=0.65).
Blind participants
Congenitally blind participants also rated protagonists in positive stories as feeling
significantly better than those in negative stories, with no effect of modality and no
interaction (hearing-bad: 1.81±0.15, hearing-good: 3.51±0.13, seeing-bad: 1.95±0.25,
seeing-good: 3.43±0.11; main effect of valence: F(1,8)=72, p<0.001, partial η2=0.72;
modality: F(1,8)=0.02, p=0.9, partial η2 <0.01; modality by valence interaction:
F(1,8)=0.5, p=0.5, partial η2<0.02).
In reaction time of the blind adults, there were no significant effects of modality or
valence, and no interaction (hearing-bad: 5.51±0.35, hearing-good: 5.02±0.41, seeingbad: 5.47±0.33, seeing-good: 5.64±0.3; all F<3.1, all p>0.1).
Group Comparison
We found no differences in the ratings across groups. We performed a 2x2x2 repeated
measures ANOVA crossing valence and modality as within-subjects factors with group
(blind/sighted) as a between-subjects factor. All participants rated the protagonist as
feeling worse in the negative valence stories (F(1,20)=264, p<0.001, partial η2=0.76),
with no main effect of modality or group, and no interactions (all F<1, p>0.3).
In the reaction time data, there were no main effects of modality, valence, or group (all
F<1, p>0.4). We found a small but significant modality by valence interaction
(F(1,20)=7.0, p=0.02, partial η2=0.03): participants responded more slowly to stories
82 of 179
about negative experiences based on hearing than seeing; and more slowly to stories
about positive experiences based on seeing than hearing.
There was also a small group by modality interaction (F(1,20)=7.4, p=0.01, partial
η2=0.02): Sighted adults respond faster than blind adults to stories about seeing; blind
adults respond faster than sighted adults to stories about hearing. Post-hoc t-tests
comparing groups within modality (e.g. RTs of blind vs sighted for seeing stories)
revealed no differences in reaction time between groups in either modality (all t<1.1, all
p>0.2).
Motion and Artifact Analysis Results
Sighted participants and blind participants showed no difference in the mean translation
per run (sighted mean + sd = 0.22mm ± 0.03, blind = 0.23mm ± 0.04), t(20)=0.13,
p=0.9), mean rotation per run (sighted = 0.26 degrees ± 0.04, blind = 0.24 ± 0.04,
t(20)=0.37, p=0.71, or the mean number of outliers per run (sighted = 0.98±0.46, blind =
0.53±0.35, t(20)=0.72, p=0.48) Together, these data suggest the groups were well
matched in motion and scanner noise.
fMRI Results: Functional Localizer
Replicating many studies using a similar functional localizer task (e.g. Saxe &
Kanwisher, 2003), we localized five theory of mind brain regions showing greater
activation for false belief stories compared to false photograph stories in the majority of
participants (uncorrected, p<0.001, k>10): Sighted participants: RTPJ 13/13
participants, LTPJ, 12/13, PC 13/13, DMPFC 11/13, VMPFC 11/13; Blind participants:
RTPJ 8/9 participants, LTPJ 9/9, PC 9/9 and DMPFC 9/9, VMPFC 9/9 (Figure 3.3 A).
As reported in (Bedny et al., 2009), sighted and blind participants did not differ in the
activation or anatomical loci of any active regions (all p>0.1), nor did blind participants
show more variability in spatial location or size of ROIs.
fMRI Results: Within ROI Pattern Analysis
Sighted participants
Source (seeing vs. hearing): Multi-voxel pattern analyses revealed reliably distinct
patterns of neural activity for stories about seeing versus hearing in the RTPJ and LTPJ,
but not PC, DMPFC, or VMPFC, of sighted adults. Note that correlations are Fisher Z
transformed to allow statistical comparisons with parametric tests. Across partitions of
the data, the pattern generated by stories in one category (seeing or hearing) was more
correlated with the pattern for the same category than with the pattern for the opposite
category, in the RTPJ (within condition correlation ± standard error, z=1.1±0.2, across
condition correlation, z=0.95±0.2, t(12)=2.0, p=0.03) and LTPJ (within = 0.82±0.2,
across = 0.64±0.2, t(11)=2.0, p=0.04), but not in the PC (within = 0.86±0.2, across =
0.78±0.2, t(12)=0.93, p=0.19), DMPFC (within=0.7±0.1, across=0.5±0.2, t(10)=1.2,
p=0.12) or VMPFC (within = 0.58±0.2, across = 0.47±0.2, t(10)=0.75, p=0.23, Figure
3.3 B, Table 3.1).
83 of 179
Figure 3.3: MVPA results. (A) Canonical theory of mind brain regions; for all analyses, theory of mind
brain regions were individual defined in each participant. (B) Sighted adults (n=13) show pattern
discrimination for beliefs based on seeing vs. hearing in the right and left TPJ, but not in any other
theory of mind region. (C) These representations of seeing and hearing persist in the RTPJs of blind
adults (n=9).
To test whether the difference between ROIs itself was significant, we conducted a 1x5
repeated measures ANOVA. Because baseline differences in correlations across ROIs
are hard to interpret (higher overall correlation in one region compared to another could
be due to many factors, including the distance of a region from the coils, amount of
vascularization, or region size; see Smith, Kosillo, & Williams, 2011), we used difference
84 of 179
scores (within condition correlation - across condition correlation) as the dependent
variable. We found a main effect of ROI (F(4,28)=3.03, p=0.03, partial η2=0.27),
suggesting that the regions contain varying amounts of information about source
modality in their neural pattern.
Valence (good vs. bad): In no region did the pattern of response distinguish between
good and bad valence (all correlation differences t<0.2, all p>0.1, Figure 3.3 B).
Blind participants
Source (seeing vs. hearing): Like in seeing adults, the pattern generated by stories
within a condition were more correlated with other stories in the same condition
compared to stories in the opposite condition in the RTPJ (within condition correlation ±
standard error, z=1.1± 0.2, across condition correlation z=0.93±0.3, t(7)=2.0, p=0.04),
but not in the LTPJ (within=1.3±0.3, across=1.3±.03, t(8)=0.2 p=0.4), PC
(within=0.71±0.2, across=0.59±0.2, t(8)=1.3, p=0.12), DMPFC (within=0.83±0.2,
across=0.73±0.1, t(8)=1.4, p=0.11) or VMPFC (within=0.56±0.2, across=0.34±0.2,
t(8)=1.4, p=0.11, Figure 3.3 C, Table 3.1). This difference between ROIs in
discrimination was significant (F(4,28) = 4.0, p=0.01, partial η2 =0.36).
Valence (good vs. bad): As in sighted adults, the pattern of response in congenially
blind adults did not distinguish between good and bad valence in any theory of mind
region (all correlation differences t<0.2, all p>0.1, Figure 3.3 C).
Sighted
Blind
Region
Within
Across
Significance
RTPJ
1.1±0.2
0.95±0.2
t(12)=2.0, p=0.03*
LTPJ
0.82±0.2
0.64±0.2
t(11)=2.0, p=0.04*
PC
0.86±0.2
0.78±0.2
t(12)=0.93, p=0.19
DMPFC
0.7±0.1
0.5±0.2
t(10)=1.2, p=0.12
VMPFC
0.58±0.2
0.47±0.2
t(10)=0.75, p=0.23
RTPJ
1.1± 0.2
0.93±0.3,
t(7)=2.0, p=0.04*
LTPJ
1.3±0.3
1.3±.03
t(8)=0.2 p=0.4
PC
0.71±0.2
0.59±0.2
t(8)=1.3, p=0.12
DMPFC
0.83±0.2
0.73±0.1
t(8)=1.4, p=0.11
VMPFC
0.56±0.2
0.34±0.2
t(8)=1.4, p=0.11
Table 3.1: Z scores for within versus across condition comparisons of seeing and hearing. * p < .05
85 of 179
Across Groups
Source (seeing vs. hearing): Comparing across groups, we looked for three things: (a)
a main effect of discrimination, driven by differences between the within-condition and
across-condition correlations in that region: evidence of reliable discrimination between
conditions; (b) a main effect of group, which indicates that one group has overall higher
correlations, due to higher overall inter-trial correlations independent of condition (and
thus not suggestive of interpretable group differences), and (c) an interaction between
discrimination and group, such that one group shows a larger difference between the
within- and across-condition correlations: evidence that one group shows more
sensitivity to condition differences than the other. Finding an interaction would be the
key piece of evidence to show that blind and sighted people have difference sensitivity
to the condition manipulation.
Overall, we found that blind and sighted adults showed very similar neural patterns, with
no evidence that either group was more sensitive to the distinction of seeing versus
hearing. We found evidence of distinct neural patterns for seeing and hearing in both
blind and sighted adults in the TPJ and no other regions. Specifically, blind and sighted
participants show equally robust neural discrimination of seeing versus hearing in the
RTPJ, with a main effect of discrimination (F(1,19) = 7.8, p=0.01, partial η2 = 0.3), no
effect of group (F(1,19)=0.001, p=0.9), and, critically, no interaction (F(1,19)=0.02,
p=0.89). There were no significant effects of group, discrimination, or their interaction in
any other ROI. LTPJ showed a trend towards both a main effect of group (overall higher
inter-trial correlations in blind participants, F(1,19)=3.2, p=0.09) and a main effect of
discrimination (F(1,19)=3.6, p=0.07), with no interaction. Precuneus, DMPFC, and
VMPFC showed no effects (PC: group: F(1,20)=0.5, p=0.5, discrimination: F(1,20)=2.3,
p=0.14, interaction: F(1,20)=0.1, p=0.7; DMPFC: group: F(1,18)=0.9, p=0.4,
discrimination: F(1,18)=2.7, p=0.12, interaction: F(1,18)=0.3, p=0.6; VMPFC: group:
F(1,18)=0.1, p=0.8, discrimination: F(1,18)=2.1, p=0.16, interaction: F(1,18)=0.3, p=0.6).
In summary, this suggests that both blind and sighted adults code the difference
between other people's experiences of seeing and hearing in the RTPJ.
Valence (good vs. bad): We found no evidence of neural code distinguishing good
valence from bad in any brain region of either group. Comparing patterns for valence,
we found a main effect of group in the LTPJ (F(1,18)=5.1, p=0.04, partial η2=0.22), due
to higher overall inter-trial correlations in the blind, independent of condition (and thus
not suggestive of sensitivity to differences in valence). There were no significant effects
of group, discrimination, or their interaction in any other ROI. "
Whole Brain Searchlight Analysis
Source (seeing vs. hearing)
The results of the searchlight converge with the ROI analyses, suggesting a
representation of seeing and hearing in the RTPJ of both groups. Because of the
similarity across groups in the ROI analyses, we combined the data from blind and
sighted participants for increased power. Converging with the results of the ROI
analyses, the whole brain analysis revealed that only the RTPJ distinguished between
86 of 179
seeing versus hearing beliefs (n = 22, peak voxel at [60, -48, 12], p<0.001, uncorrected
peak T = 3.6).
Second, a two-sample T-test across groups revealed a significant group difference in
the left dorsolateral PFC (BA45/46, peak voxel at [-58, 30, 8], p<0.001 uncorrected,
peak T = 3.7). In this DLPFC region, sighted participants showed a greater difference
between within and across condition correlations than blind participants. No regions
showed stronger decoding in blind participants relative to sighted ones.
Valence (good vs. bad):
Combining across groups for power, we find that only the anterior cerebellum
distinguishes between good and bad emotional valence (n = 22, peak voxel at [12, -42,
-32], p<0.001 uncorrected, peak T = 4.2).
87 of 179
Discussion
MVPA reveals features of mental state representations
For neuroimaging research to have cognitive implications, a key challenge is to go
beyond where a cognitive function occurs in the brain, to provide a window into neural
representations and computations. Dozens of neuroimaging studies suggest a
hypothesis for where in the brain key aspects of theory of mind are processed (e.g.,
Aichhorn et al., 2009; Bedny, Pascual-Leone, & Saxe, 2009; Lombardo, Chakrabarti,
Bullmore, Baron-Cohen, & Consortium, 2011; Mason & Just, 2011; Rabin, Gilboa,
Stuss, Mar, & Rosenbaum, 2010; Saxe & Kanwisher, 2003; Saxe & Powell, 2006;
Spunt, Satpute, & Lieberman, 2010). The right and left TPJ, the PC, and MPFC show
increased hemodynamic responses to stimuli that require participants to think about
other people’s mental states, and many studies have investigated the selectivity and
domain specificity of these brain regions for theory of mind. These brain regions
therefore provide a prime opportunity to probe the neural computations of mental state
representation: What features of people’s beliefs and intentions are represented, or
made explicit, in these brain regions?
A powerful, if simplifying, assumption about neural representation is that a feature is
explicitly represented by a population of neurons if that feature can be decoded linearly
from the population’s response. Different subpopulations of neurons within a region may
contribute to a common task by representing different features or aspects of the
stimulus or task. A well-studied example is object recognition in the ventral pathway of
macaques. Low-level features of the stimulus, like retinotopic position and motion
energy, can be linearly decoded from the population response in early visual cortex,
whereas higher-level properties, like object identity (invariant to size and position), can
be linearly decoded from the population response in IT, a region further along the
processing stream (DiCarlo et al. 2012; Kamitani & Tong 2005).
We can take advantage of this property of neural populations to identify features of
human neural representations using fMRI. When the subpopulations of neurons that
respond to different stimulus features are at least partially organized into spatial clusters
or maps over cortex (Dehaene & Cohen, 2007; Formisano et al., 2003), those features
may be detectable in reliable spatial patterns of activity measurable with fMRI (Haynes
& Rees, 2006; Kriegeskorte & Bandettini, 2007; Norman et al., 2006). Multi-voxel
pattern analyses therefore offer an exciting opportunity for studying the representations
underlying human theory of mind. In spite of many studies on the neural basis of theory
of mind, very little is known about which features or dimensions of mental states are
represented by the neural populations in the candidate brain regions. Here we identified
two dimensions of others’ mental states that are common and salient: the source
modality of the other person’s experience, and the emotional valence. To ask whether
either of these dimensions are explicitly represented in these brain regions, we then
tested whether either of these dimensions could be linearly decoded from the spatial
pattern of the response in any theory of mind brain region.
We found that both the right and left TPJ of sighted people showed distinct spatial
patterns of responses to stories that described seeing versus hearing. Two independent
88 of 179
groups of stories, similar only in the source modality of the character’s beliefs, elicited
reliably similar spatial patterns of response in the TPJ, in spite of the wide heterogeneity
of the stories in many other respects (e.g. the physical environment, the type of event,
the character’s name, age, gender, social status, etc). Furthermore, we replicated these
results in the right TPJ in an independent group of congenitally blind adults. We
therefore suggest that spatially distinguishable neural populations in the TPJ explicitly
represent the perceptual source of another person’s mental states. Together with prior
evidence, our findings suggest that these representations develop in the absence of first
person experience.
Knowledge source and theory of mind
Why might the TPJ code the distinction between seeing and hearing mental states
described in stories? One possibility is that the patterns of activation we observed were
driven simply by a response to the mental state attitude verb used in the story (i.e.
“sees” vs “hears”). However, we consider this interpretation unlikely, for three reasons.
First, mental state verbs on their own elicit little to no response in the TPJ, both relative
to rest and relative to non-mental verbs (e.g. “to think” vs “to run” or “to rust,” Bedny,
McGill, & Thompson-Schill, 2008). Robust responses are best elicited in TPJ by full
propositional attitudes (e.g., a person’s mental attitude towards some content), ideally
embedded in an ongoing narrative. Second, in previous research we directly
manipulated the mental state verb in a sentence and did not find distinct patterns of
response in TPJ (e.g. “believe” vs “hope”, Paunov and Saxe, personal communication).
Third, in an other experiment, we found distinct patterns of response in TPJ for distinct
mental states that were described using the same verb (i.e. intentional vs accidental
harm, both described by “believed”, Chapter 2, Koster-Hale et al., 2013). In all, we
suggest that the patterns of activity observed here are most likely responses to features
of the character’s mental state, rather than to specific verbs in the stories. Still, future
research should test this possibility by independently manipulating the source of the
mental state and the verb used to describe it, as well as by looking at the pattern
evoked by these verbs outside of a mental state context.
Based on the current findings, we hypothesize that the TPJ contains an explicit
representation of the perceptual source of mental states. Why might theory of mind
brain regions code this information? Considering the source of others’ knowledge is
central to inferences about other people’s minds. Knowing how another person got their
information can influence the inferences we make about what they know, how well they
know it, and whether we should believe it ourselves. For example, we are more likely to
trust another person’s testimony if they acquired their knowledge by direct visual
observation (an “eye-witness”) than through gossip or hearsay (Bull Kovera, Park, &
Penrod, 1992; Peter Miene, Bordiga, & Park, 1993; Peter Miene, Park, & Bordiga,
1992); eyewitness but not hearsay testimony is admitted as evidence in courts of law
(Park, 1987). Similarly, recalling how one acquired one’s own information plays an
important role in evaluating and justifying a belief, and deciding how readily it should be
discarded (O'neill & Gopnik, 1991).
89 of 179
Recognizing sources of knowledge is so cognitively salient that some languages
explicitly mark source syntactically. In languages with evidential marking, such as
Turkish, Bulgarian, Tibetan and Quechua, utterances are syntactically marked to
indicate how the information was acquired (Aikhenvald, 2004; J. G. de Villiers, Garfield,
Gernet-Girard, Roeper, & Speas, 2009; Faller, 2002; Fitneva, 2001; Johanson & Utas,
2000; Smirnova, 2013). Some languages mark direct versus indirect experience, or
hearsay versus everything else; others mark whether the information was gained from
inference, hearsay, or sensory experience. Languages such as Tucano and Fasu
encode aspects of sensory modality, including distinctions between visual versus nonvisual sensory information; one language, Kashaya, has a separate marker for auditory
evidence (Aikhenvald, 2004). Acquisition of linguistic evidential marking by children
appears to depend on the development of theory of mind: during acquisition of Turkish,
for example, children’s earlier performance on explicit theory of mind tasks concerning
sources of knowledge predict their later comprehension of linguistic evidentials, but not
vice versa (Papafragou & Li, 2001).
In sum, the source of others’ knowledge is both behaviorally relevant and cognitively
salient; the current results suggest that it is also explicitly encoded by distinct neural
populations within the TPJ. An open question for future work is whether patterns in the
TPJ can distinguish only the modality or also other aspects of another person’s source
of knowledge. For example, sources of knowledge can be direct (perceptual
observation), inferential (by induction from clues or patterns), or from hearsay (from
other people’s report). Direct perceptual sources may be more or less reliable,
depending on the situation (e.g. observed from nearby or at a distance, based on a
quick glance or a long stare) and the observer (e.g. an expert, an amateur, or a child).
Future research should investigate the neural representation of these other features of
knowledge source, and the relationship between these features and perceptual source.
The valence of others’ mental states
In contrast to perceptual source, we failed to decode the valence of another person’s
mental experience from the pattern of activity in any theory of mind region. This is
especially noteworthy, considering that the behavioral task — judging how good or bad
someone felt at the end of the story — specifically drew attention towards valence, while
ignoring source information. However, a prior study successfully decoded positive
versus negative valence of others’ emotions, expressed in emotional body movements,
facial expressions, and tones of voice, from patterns of activity in MPFC (Peelen,
Atkinson, & Vuilleumier, 2010; Skerry & Saxe, in prep). One possibility is that emotional
experience is more effectively conveyed in non-verbal stimuli. However, null results from
MVPA should be interpreted with caution. The neural sub-populations that respond to
different types of valence could be intermingled such that we would be unable to detect
the distinct neural populations at the spatial scale of fMRI voxels; the distinction would
only be revealed by techniques with higher spatial resolution.
90 of 179
Blind adults represent knowledge source
We found similar patterns of neural activity in blind and sighted adults. Most notably, we
found that other people’s visual versus auditory experiences are encoded distinctly by
both sighted and blind adults. Blind individuals specifically lack first person experience
of visual aspects of mental states, and thus allow us to test the role of first person,
sensory experience in the development of theory of mind representation: what does it
take to be able to think about someone else’s experience?
One set of theories of how we understand others’ mental experiences posits that we
vividly simulate, or imagine, having the same experience ourselves. For example,
Gallese & Sinigaglia (2011) describe the process of simulation as: “People reus[ing]
their own mental states or processes in functionally attributing them to others, where the
extent and reliability of such reuse and functional attribution depend on the simulator’s
bodily resources and their being shared with the target’s bodily resources.” (p. 518).
That is, we understand other people’s mental states using a template of our own firstperson experiences. This view is difficult to reconcile with evidence that blind people
use similar neural mechanisms to reason about seeing as do sighted people. In the
initial analyses of these data, Bedny et al. (2009) showed that blind adults can represent
mental states acquired by vision, without additional costs or compensatory
mechanisms. The present analyses further show that in these same blind and sighted
adults, not only are the same brain regions recruited to think about others’ mental states
based on seeing and hearing, but these regions represent the difference between visual
and auditory sources of belief. These results suggest the representation of perceptual
source in the TPJ is not dependent on first-person experience.
Importantly, however, our data do not show that people never use first-person
experience to make inferences about other people’s minds and goals. Both in the motor
and sensory domains, people do seem to use sensory-motor information to understand
other’s actions and experiences. For example, interference between other people’s
sensory-motor experience and our own suggests that we use partially shared
representations for action observation and action execution. Observing someone else’s
actions can interfere with executing actions yourself (Kilner, Paulignan, & Blakemore,
2003), and observing what another person can see interferes with the ability to report on
what you yourself are seeing (Samson, Apperly, Braithwaite, Andrews, & Bodley Scott,
2010). Similarly, personal experience specifically helps children learn about others’
visual experiences: infants who experienced a transparent blindfold follow the gaze of a
blindfolded adult, but infants who experienced an opaque blindfold do not (Meltzoff,
2004; 2007; Meltzoff & Brooks, 2007).
In all, understanding of other minds is likely guided by both first-person experience,
which provides rich and detailed input about the character of specific experiences, and
an intuitive theory of mind, which allows individuals to form representations of
experiences they have not had, and indeed could never have, such as blind adults’
representations of experiences of sight. In some cases, such as action prediction, first
person sensory machinery may play a direct role in representing information about other
people; in other cases, first person experience provides a rich source of data about how
people feel and act, playing an important role in building causal theories of other’s
91 of 179
minds. The best “theory of mind” will both be able to deal with abstract generalizations
and make use of data from first-person experience (Apperly, 2008; Keysers & Gazzola,
2007; Kilner & Frith, 2007).
Learning about sight
The present findings, along with prior evidence, suggest that blind people can have
impressively rich knowledge about sight (Landau & Gleitman, 1985; Marmor, n.d.;
Shepard & Cooper, 1992). How do blind individuals acquire this extensive knowledge
about seeing? We propose that one source of information is others’ testimony. Blind
children live immersed in a world of sighted people, who talk constantly about
experiences of seeing. In general, children acquire much of their knowledge of the
invisible causal structure of the world through testimony (Harris, 2002a; 2002b; Harris &
Koenig, 2006; Koenig, Clément, & Harris, 2004). Language and testimony are a
particularly powerful tool for learning about the invisible contents of other minds (Bedny
& Saxe, 2012; Robinson & Whitcombe, 2003; Robinson, Haigh, & Nurmsoo, 2008;
Urwin, 1983); the absence of linguistic access to other minds can significantly delay
theory of mind development, for example in deaf children born to non-signing parents
(Figueras-Costa & Harris, 2001; Meristo et al., 2007; de Villiers, 2005; Moeller & Schick,
2006; Peterson & Wellman, 2010; Peterson & Siegal, 1999; Pyers & Senghas, 2009;
Schick, de Villiers, de Villiers, & Hoffmeister, 2007; Woolfe & Want, 2002). In contrast,
conversational access seems to give blind children a detailed representation of visual
mental states, and the inferences they afford.
The importance of testimony to the understanding of other minds is highlighted by the
contrast between how much blind people know about sight and how little sighted people
know about blindness. Blind individuals can never directly experience vision, but they
are constantly exposed to testimony about seeing. By contrast, sighted people do
experience the perceptual aspects of blindness (when in the dark or with their eyes
closed). Nevertheless, the average sighted person has little understanding of blindness.
Most sighted people do not know how blind individuals navigate their environment or
perform daily tasks such as dressing and eating. This lack of knowledge can lead to
incorrect and sometimes problematic inferences about blind people’s incapability to
work, learn, and parent. Given how much blind people know about seeing, such
misunderstandings are unlikely to result from an in principle incapacity to understand
experiences different from our own. While blind people are surrounded by sighted
people, most sighted individuals have never met anyone who is blind. The ignorance of
sighted people is thus another dramatic example of the importance of testimony for
learning about other minds.
However, these experiments have just begun to probe sight understanding in blind
individuals. In the current stimuli (and unlike our initial example of the woman at the
window) neither the content nor the reliability of the mental states depended on the
perceptual modality. The belief content (e.g. “his team lost”) was matched across
perceptual sources, and always reliably inferred from the perceptual evidence. Thus,
successfully encoding the story did not depend on specific knowledge about sight, e.g.,
how difficult it is to recognize a face in the dark. It is possible that blind people would
92 of 179
show different behavior and neural processes for such inferences. Additionally,
participants’ task (judging how good or bad the protagonist would feel) depended on
encoding the content but not the source of beliefs. Despite this, participants
spontaneously encoded the perceptual source of the belief described in the story.
However, as a consequence, the current results cannot determine whether blind and
sighted people would show similar neural patterns (or behavioral performance) if
specific details about the perceptual source were more relevant in the task. Future
research should investigate whether blind and sighted adults have similar knowledge
and representations of the details of sight, such as how knowledge based on vision will
vary with distance, occlusion, and darkness, and whether the neural patterns observed
here encode similar information about seeing and hearing. Such specific details may
show greater influences of first person experiences.
Finally, while the data here suggest a common endpoint in the development of an adult
theory of mind, the processes by which blind and sighted children acquire this theory
may be different. Although blind adults appear to have typical theory of mind, blind
children are delayed on a variety of theory of mind tasks, including both tasks that rely
directly on inferences about vision, such as perspective taking tasks (Bigelow, 1991;
1992; Millar, 1976), and also tasks that do not depend directly on visual perspective,
including auditory versions of the false belief task (McAlpine & Moore, 1995; Minter, &
Hobson, 1998; Peterson et al., 2000). Using a battery of theory of mind tasks designed
specifically for blind children, Brambring & Asbrock (2010) concluded that blind children
are delayed an average of 1-2 years, relative to blindfolded sighted control children,
though these children catch up in adolescence (see Peterson & Wellman, 2010 for a
similar argument about deaf children). Thus an open question is whether the neural
mechanisms for theory of mind, which are similar in blind and sighted adults, develop
differently in blind and sighted children.
Conclusion
In summary, we find that (i) theory of mind brain regions (specifically, the TPJ) encode
perceptual source and sensory modality of others’ mental states, and (ii) these
representations are preserved in the RTPJ of congenitally blind adults. Considerable
neuroimaging work on theory of mind suggests that the RTPJ plays a role in thinking
about others’ thoughts; we find that one aspect of this role is to make explicit, in the
population response of its neurons, features of beliefs — in this case, the perceptual
source of the belief. The persistence of these representations in congenitally blind
adults, provides evidence that theory of mind brain regions come to encode these
aspects of mental states even in the absence of first-person experience.
93 of 179
Acknowledgments
We thank the Athinoula A. Martinos Imaging Center at the McGovern Institute for Brain
Research at MIT, and the Saxelab, especially Amy Skerry and Hilary Richardson for
helpful discussion. We also gratefully acknowledge support of this project by an NSF
Graduate Research Fellowship (#0645960 to JKH) and an NSF CAREER award
(#095518), NIH (1R01 MH096914-01A1), and the Packard Foundation (to RS).
94 of 179
Chapter 4.
Mentalizing regions represent continuous, abstract
dimensions of others’ beliefs 10
The human capacity to think about others’ thoughts includes not only the capacity to
infer what another person is thinking, but critically also the ability to evaluate why the
other person believes what they do. In two experiments, we investigate how the
sources of others’ beliefs are coded in brain regions involved in mental state
inference, including bilateral temporo-parietal junction (TPJ), superior temporal
sulcus (STS) and medial prefrontal cortex (MPFC). We probed three epistemic
features of others’ beliefs (i.e. distinctions relevant to the source of another person’s
belief), and one salient but non-epistemic distinction (the valence of the person's
resulting emotion). Using multivoxel pattern analyses (MVPA), we find that the
patterns of activity in right TPJ and right STS contain complementary neural codes,
which reliably distinguish stories in which a person comes to a conclusion based on
reliable and unambiguous evidence, from stories in which they come to their
conclusion based on unreliable and ambiguous evidence. The neural classification
accuracy for each story was predicted by independent ratings of the strength of the
evidence in that story. We also find evidence for a representation of
evidence modality (visual vs auditory) in bilateral TPJ, and of emotional valence in
ventral MPFC. Together, these data suggest that the cortical regions implicated in
mental state inference contain neural representations of epistemic and emotional
features of others’s beliefs, and that these features are represented along
continuous, abstract dimensions. 10
Koster-Hale, J., Richardson, H., Velez-Alicia, N, Asaba, M, Young, L., & Saxe, R. R. (in prep). Thinking
about seeing: Mentalizing regions represent continuous, abstract dimensions of others’ beliefs
95 of 179
Introduction
Othello is a tragedy not only because Othello is wrong about Desdemona — she is not
unfaithful to him — but because he is wrong for the wrong reasons. Seeing a lost
handkerchief, hearing insinuations from a jealous and vengeful subordinate: Othello has
only poor, indirect, and unreliable evidence against his wife. And yet, as Iago tells the
audience, “trifles light as air/ are to the jealous confirmations strong/ as proofs of holy
writ.” (Othello, 3.3.322-324). For the audience, it is almost unbearable to watch Othello
allowing himself to believe, and to act on, such trifles.
Understanding why someone believes what they do — their evidence and reasons for
believing — is a key component of human social cognition. Observers readily make
inferences about the source of other people’s beliefs (Does the evidence really justify
their conclusion? Did they learn the information themselves, or did someone tell them?).
We can use information about about the source of others’ beliefs to facilitate learning
and cooperation (Robinson & Whitcombe, 2003; Smith & Ellsworth, 1987; Toglia, Ross,
Ceci, & Hembrooke, 1992; Whitcombe & Robinson, 2000); to judge and to punish
(Young, Nichols, & Saxe, 2010), and to decide who to trust in the future (Jaswal &
Neely, 2006; Koenig & Harris, 2005a; Scofield & Behrend, 2008). We more readily
believe people with knowledge from direct visual observation (an “eye-witness”) than
from inference or hearsay (i.e. other people’s report; Bull Kovera, Park, & Penrod, 1992;
Peter Miene, Bordiga, & Park, 1993; Peter Miene, Park, & Bordiga, 1992; Robinson,
Champion, & Mitchell, 1999) and we forgive those who made justifiable mistakes, based
on the evidence available to them, even when they cause harm (Kornblith, 1983; Young
et al., 2010). Yet, despite its prominent role in human social cognition, little is known
about how the brain tracks and represents the relationship that other people have to the
evidence available in the world.
Previous work suggests a set of cortical regions that may represent abstract features of
others’ minds. Regions implicated in mental state inference, including bilateral temporal
parietal junction (right and left TPJ), right superior temporal sulcus (STS), precuneus
(PC), and medial prefrontal cortex (MPFC), collectively called the “theory of mind (ToM)
network,” show selective activity while participants are thinking about other people and
especially their minds, and these effects are robust across stimuli, tasks, and
populations (for a review see Chapter 1, and Koster-Hale & Saxe, 2013). There is thus
extensive evidence concerning when, how much, and how selectively these regions are
recruited when thinking about other minds. Yet, relatively little is know about the
representations and computations supported by the populations of neurons in these
regions.
A powerful approach for understanding abstract neural representation is to ask which
features of a stimulus can be decoded from a population of neurons. Many high-level,
abstract features are coded in the pattern of response in a population of neurons (e.g.
Averbeck, Latham, & Pouget, 2006; DiCarlo, Zoccolan, & Rust, 2012; Quian Quiroga &
Panzeri, 2009; Tudusciuc & Nieder, 2007). Though fMRI data pools over hundreds of
thousands of neurons, if these subpopulations of neurons are spatially organized into
clusters or maps over cortex (Dehaene & Cohen, 2007; Formisano et al., 2003; Konkle
96 of 179
& Caramazza, 2013), then these representations may be detectable with fMRI, using a
class of analysis techniques called multi-voxel pattern analyses (MVPA; Haynes &
Rees, 2006; Kriegeskorte, Bandettini, & Bandettini, 2007; Norman, Polyn, Detre, &
Haxby, 2006). MVPA has been successfully used to probe the neural basis of many
different types of representation, including emotional affect and person identity
(Anzellotti, Fairhall, & Caramazza, 2013; Chikazoe, Lee, Kriegeskorte, & Anderson,
2014). Support vector machines (SVMs) are a particularly powerful version of linear
classifiers, because their relatively few free parameters and robustness against
overfitting (Hearst, Dumais, Osman, Platt, & Scholkopf, 1998; Steinwart & Christmann,
2008; V. N. Vapnik & Vapnik, 1998), and can yield trial-by-trial information, allowing us
to relate features of individual stimuli to the neural response they generate. Here, we
use these methods to probe the neural representations of others’ beliefs.
Specifically, we ask whether regions implicated in social cognition contain neural
populations that track features of others’ beliefs, including three epistemic distinctions
(i.e. those that are relevant to the source of others’ beliefs), and contrast this with one
salient but non-epistemic distinction: the valence of the person’s resulting emotion. We
independently manipulated three features of other people’s evidence: First, and mostly
critically, the quality of other people’s evidence: how good or bad was the evidence, for
the conclusion? Second, the perceptual modality of the evidence: was the other
person’s evidence seen or heard? Third, the directness of the evidence: was the other
person's evidence experienced first-hand or learned indirectly from another person’s
report? Finally, for comparison, we included a non-epistemic feature of other people’s
mental states: whether the resulting emotion is positive or negative (emotional valence).
To answer the question of how these features are represented, we first examined
whether population responses within theory of mind regions could reliably discriminate
between categories: good versus bad evidence; visual versus auditory evidence; firstperson experience versus hearsay, and positive versus negative emotion. We then
examined whether these representations were sensitive to continuous measures of
these dimensions that vary across individual stimuli. Lastly, we examined whether these
representations replicated and extended in a second experiment, manipulating the
quality of others’ evidence in the context of moral judgments.
97 of 179
Methods
Participants
Twenty right-handed members of the larger MIT community participated (11 women;
mean age ± SD, 28.5 years ± 6.8). All participants were native English speakers, had
normal or corrected-to-normal hearing and vision, gave written informed consent in
accordance with the requirements of Institutional Review Board at MIT, and received
payment. Two participants were dropped for excessive head motion, and one failed to
complete the experiment, leaving 17 in all analyses.
Stimuli
Experiment 1 featured 50 stories in 5 conditions. 10 stories were control stories,
describing physical events and scenes (such as gardens full of colorful flowers). In the
remaining 40 target stories, the protagonist came to hold a belief about the world, based
on one of four different types of evidence (i.e. four epistemic conditions). In Seeing
Good Evidence (SG) stories, the protagonist makes an inference based on clear visual
access to reliable and unambiguous evidence (e.g., a woman infers that a wolf is
outside, based on seeing large, fresh paw-prints, in good light). In Seeing Bad
Evidence (SB) stories, the protagonist makes the same inference based on indistinct
and unreliable visual evidence (e.g., the woman sees some old markings in the dirt in
dim light). In Hearing First-person Evidence (HF) stories, the protagonist draws her
conclusion based on good aural evidence (e.g., the woman hears a distinctive growl on
a quiet afternoon). Finally, in Hearsay Evidence (HS) stories, the protagonist comes to
her conclusion based on someone else’s report of evidence (e.g., the woman is told by
her friend that there are large, fresh paw prints in the dirt). In half of the stories in each
condition, the protagonist felt a negative emotion based on her inference (e.g. that
there is wolf outside), whereas in the other half of the stories, the protagonist would feel
a positive emotion based on her inference (e.g. that she is getting a puppy for her
birthday, Figure 4.1).
Evidence quality was manipulated in two ways. First, we changed the characters’
perceptual access to the evidence. In Seeing Good-Evidence stories, protagonists saw
evidence from close up, in bright lighting, and in still, visually clean environments, while
in Seeing Bad-Evidence stories, protagonists saw evidence through visual barriers
(such as dirty windows, or foggy air), in poor lighting, from far away, and in cluttered
environments. Second, we changed the extent to which the evidence supported the
conclusion. SG stories contained evidence that relatively directly and unambiguously
supported the protagonist’s conclusion (such as a pirate inferring his treasure has been
stolen after seeing a large hole and shovels where he buried his treasure), while in SB
stories the evidence was more ambiguous (the pirate reaching the same conclusion
after seeing a ship sailing off on the horizon line before he gets to his treasure site). The
bad evidence could both be visually ambiguous (concluding that a good friend has
arrived, from seeing only the outline of a person, compared to directly seeing his face)
and conceptually ambiguous (concluding that your favorite pen was stolen after seeing
two bullies give each other high-fives, compared to seeing them pocket something silver
98 of 179
and pen-sized). Directness was always the difference between the protagonist hearing
direct evidence with their own ears, versus being told about the evidence. The reported
evidence was always the same evidence presented in the SG or HF condition. Thus,
even in the HS stories, the protagonist is drawing their own conclusion based on the
evidence they were told about. Finally, modality was manipulated by changing whether
the protagonist received the information visually or aurally.
Individual participants saw each target story in only one of the four epistemic conditions,
forming a matched and counterbalanced design; every target story occurred in all four
conditions across participants. The valence of the protagonist’s emotion was
manipulated across stories; half of the stories in each condition led to a positive
emotion, and half led to a negative emotion.
Seeing
Good*Evidence*(SG)
Background*
(16s)
Evidence
(4?16s)
Seeing
Bad*Evidence*(SB)
Hearing*
First?person*(HF)
Hearing*
Hearsay*(HS)
Julie'had'wanted'a'puppy'of'her'very'own'for'years.'Every'holiday,'she'asked'her'parents'to'give'
her'a'puppy,'but'they'had'always'said'that'they'couldn’t'afford'it.'On'Julie’s'ninth'birthday,'she'
woke'up'and'ran'downstairs'to'her'parents.
In'the'cluBered'In'the'bright-kitchen,'
kitchen,'she'saw'a'
she'saw'a'newlyfew-small-toys-on-theinstalled-doggy-door.'
floor.
In'the'quiet'kitchen,'
she'heard'a'quiet,excited-yipping.'
Conclusion
(2?4s)
Julie'asked'her'mom,'"Did'we'get'a'new'puppy?!"
Task
Happy'or'Sad?
In'the'kitchen,'hermom-told-her'that'
there'was'a'newlyinstalled-doggy-door.'
Figure 4.1: Example stimuli from Experiment 1. In the fMRI task, participants listened to audio
recordings of 40 target stories and 10 physical control stories. Stories lasted from 22 to 36 seconds.
After each story, participants heard the question “Happy or sad?”, and indicated whether the main
character in the story felt happy or sad, using a button press (left/right).
Online Behavioral and Language Norming
To ensure that the condition manipulations were effective, we piloted 50 possible target
stories, each appearing in all four epistemic conditions (SG, SB, HF, HS; 200 items
total). Stories were rated by 521 adults in an online behavioral study using Amazon’s
Mechanical Turk (www.mturk.com), an online crowd-sourcing website that has been
found to yield valid and reliable data (Buhrmester, Kwang, & Gosling, 2011). All adults
were from the United States, and gave informed consent in accordance with the
requirements of the Institutional Review Board at MIT.
At the start of the norming study, participants were shown two example stories and
answers to ensure comprehension of the task. Participants then read the target story.
After reading each story, participants answered six questions. First, we verified that
participants could infer the modality of the evidence, by asking “At the end of the story,
[character] concludes that [conclusion]. Why did they think that?” Participants had a
99 of 179
forced choice response of “Something they heard” and “Something they saw”. Second,
we asked about the quality of the evidence, by asking “At the end of the story,
[character] concludes that [conclusion]. How good was his/her evidence?” (Scale 1 (not
good at all) to 7 (very good)). Third, we asked “How engaging was the story?” (1 (not at
all) to 7 (very engaging)). Fourth, we asked “How easy was it to read the story?” (1 (not
at all) to 7 (very easy)). Fifth, because the explicit task for fMRI participants was to rate
the character’s emotions, we asked “How happy does character feel at the end of the
story?” 1 (not at all happy) to 7 (very happy). Finally, we included a catch question,
asking “What is the name of the main character in the story?”
The study was designed so that participants never rated the same story in multiple
conditions. Four participants were excluded for (a) failing to follow basic instructions, (b)
leaving more than 50 percent of questions blank, or (c) failing 50 percent of the catch
questions, across items. Additionally, 182 ratings on single items were excluded
because (a) more than 50% of questions were left blank or (b) the catch question was
not answered correctly, leaving 521 participants who generated 3938 ratings.
Participants rated an average of 7.6 items, split evenly across conditions (SG: 2.36±1.9;
SB: 2.44±1.9; HF: 2.44±1.9; HS: 2.43±1.9). Each item was rated by an average of 19.7
people (16-21 people), split evenly across conditions (SG: 19.66±1.08; SB: 19.6±1.05;
HF: 19.8±0.88; HS: 19.7±1.02). Item-wise means were calculated as the average of all participant ratings for each item,
in each condition. Based on the behavioral ratings, 40 of the 50 target stories were
included in the fMRI experiment. These stories were rated as easy and engaging to
read, and were selected to maximize the difference in rated evidence quality between
SG and SB, and preserve the match in evidence quality between seeing and hearing,
and HF and HS, while keeping valence ratings, word count, reading ease, and
engagingness matched between conditions (Table 4.1).
Additionally, to match items in their reading ease and lexical frequency, we ran all
stories through the Coh-Metrix, an online tool that calculates a number of metrics
designed to measure computational cohesion and text difficulty (http://cohmetrix.com/’;
Graesser & McNamara, 2011; McNamara, Graesser, McCarthy, & Cai, 2014). For each
stimulus, we analyzed the Flesch reading ease (Flesch, 1948), Flesch-Kincaid Grade
Level (Kincaid, Fishburne, Rogers, & Chissom, 1975), and total word count and log
lexical frequency, both of which affect comprehension and reading time (Rayner & Duffy,
1986).
The behavioral ratings verified that our manipulation of evidence quality and valence
was effective: Seeing stories with Good evidence were rated as having better quality
evidence than Seeing stories with Bad evidence (t(77.2)=7.15, p<0.001, d=1.6 (large)).
In contrast, there was no difference in the quality of evidence between hearing firstperson and hearsay (t(77.2)=0.94, p=0.35), nor between seeing and hearing
100 of 179
(t(149.4)=-1.05, p=0.29)11 . The four epistemic conditions were matched on emotional
valence (F(3,156)=0.03, p=0.99). The protagonists of positive stories were rated as
significantly happier than the protagonists of negative stories (Positive: 6.13±0.06;
Negative: 1.69±0.03; t(110.5)=63.35, p<0.001, d=10.02 (large)).
Seeing Good
Evidence
Seeing Bad
Evidence
Hearing First
Person
Hearing
Hearsay
Evidence Quality
5.21±0.13
3.98±0.12
4.82±0.13
4.66±0.12
Valence
3.97±0.36
3.83±0.35
3.91±0.37
3.95±0.36
Engagingness
4.96±0.06
4.78±0.08
4.96±0.07
4.78±0.06
Reading Ease
6.31±0.03
6.23±0.04
6.28±0.04
6.16±0.04
Flesch Reading Ease
83.0 ± 6.6
82.5 ± 6.1
81.4 ± 6.8
82.5 ± 6.8
Flesch-Kincaid Grade Level
4.3 ± 1.2
4.4 ± 1.1
4.5 ± 1.1
4.5 ± 1.2
Word Count
68.2 ± 4.6
68.5 ± 4.4
67.5 ± 4.4
68.7 ± 4
Log Lexical Frequency
2.93 ± 0.13
2.95 ± 0.12
2.98 ± 0.13
3.01 ± 0.14
Table 4.1: descriptive statistics of fMRI stimuli used in Experiment 1. 521 participants on Amazon’s
Mechanical turk rated evidence quality, emotional valence, engagingness, and reading ease
on a scale from 1-7. Flesch Reading ease (range of 0 to 120), Flesch-Kincaid Grade Level (range of 1
to 12), word count, and log lexical frequency were calculated using Coh-Metrix, an online tool for
analyzing text.
Key Comparisons
This design allows us to test four key comparisons: First, by comparing Seeing GoodEvidence stories and Seeing Bad-Evidence stories, we can look for sensitivity to the
quality of others’ evidence (good versus bad), holding constant the modality (seeing)
and the directness (first-person). Second, by comparing Hearing First-Person stories to
Hearsay stories, we can look for sensitivity to the directness of others’ evidence (firstperson versus third-person), while holding constant the quality of the evidence and the
modality. Third, by comparing the two seeing stories (SG and SB) to the two hearing
stories (HF and HS), we can look for sensitivity to the modality of others’ evidence
(hearing vs. seeing) while holding constant the quality of the evidence (see evidence
from behavioral norming). Finally, by comparing positive to negative stories, we can look
for sensitivity to the valence of others’ emotions.
11
Note that there was a reliable difference in the quality of evidence between the Seeing Good-Evidence
stories and the First-person Hearing stories (SG: 5.21±0.13; HF: 4.82±0.13; t(78)=2.14, p=0.04, d=0.48
(small)). As a result, this specific comparison is confounded between modality and quality. All analyses of
the effect of modality use the full set of seeing and hearing stories, which are matched on evidence
quality.
101 of 179
fMRI Task
Participants were scanned while listening to audio recordings of 40 target stories and 10
physical control stories. All target stories had 16 seconds of background information
(identical across condition, 32 words ± 2.5), 6-14 seconds of evidence-gathering (in four
conditions; 25 words ± 2.3; matched across conditions; SG: 25.2±2.5, HF: 24.5±2.2, SB:
25.4±2.4, HS: 25.6±2.3), 2-4 seconds of conclusion (identical across condition, 10
words ± 2.4), and 6 seconds of response time (Figure 4.1). Stories were presented in a
pseudorandom order, condition order was counterbalanced across runs and subjects,
and no condition was immediately repeated. Rest blocks of 12 seconds occurred four
times in each run. The total experiment of five runs (6.2 minutes each) lasted 31
minutes. The stimuli were digitally recorded by 11 female speakers at a sampling rate of
44,100 to produce 32-bit digital sound files. During the stories, one of eight colorful
abstract images was displayed.
After each story, participants heard a recording of a male speaker ask “Happy or sad?”
Participants indicated whether the main character in the story felt happy or sad, using a
button press (left/right). During the response period, a screen with a happy (left) and
sad (right) face appeared, to remind participants which button to press. Reaction time
was measured from the onset of the question “Happy or sad?”. In each condition, 5
stories were happy (the protagonist felt good at the end of the story) and 5 stories were
sad.
Theory of Mind Localizer
Participants in both experiments were also scanned on a localizer task designed for
identifying theory of mind brain regions in individual participants (Dodell-Feder, KosterHale, Bedny, & Saxe, 2011; Saxe & Kanwisher, 2003); stimuli are available at http://
saxelab.mit.edu/superloc.php. Participants read verbal narratives in two conditions:
stories that described a situation in which someone held a false belief about the world,
such as mistaking white paint for cream or not knowing that her car had been moved
(belief condition); and stories describing a situation in which there was a false physical
representation of the world, such as an out-of-date photograph or advertisement
(physical condition). Both kinds of stories require the reader to consider incorrect or
outdated representations of the world, and so are similar in their meta-representational
and logical complexity; however, they differ in whether the reader is considering
someone else’s mental state, and thus comparing them serves to localize regions
recruited during mental state inference. Stories were presented in white, 24-point font
on a black background and were presented for 10s, followed by a true or false question
(4s) about the story. Each run (4.53min) consisted of 10 trials interleaved with 12s rest
periods. Participants saw two runs of the localizer (10 trials of each condition); The
orders of stimulus type (Belief or Photo) and correct answer (True or False) were
counterbalanced within and across runs.
Acquisition and Preprocessing
fMRI data were collected in a 3T Siemens scanner at the Athinoula A. Martinos Imaging
Center at the McGovern Institute for Brain Research at MIT, using a 32-channel phased
102 of 179
array head coil. We collected a high-resolution (1mm isotropic) T-1 weighted MPRAGE
anatomical scan, followed by functional images acquired with a gradient-echo EPI
sequence sensitive to blood-oxygen-level-dependent (BOLD) contrast (repetition time
(TR)=2s, echo time (TE)=40ms, flip angle=90°, voxel size 3x3x3mm, matrix 64x64, 26
near-axial slices). Slices were aligned with the anterior/posterior commissure and
provided whole-brain coverage (excluding the cerebellum). For steady state
magnetization, the first 4 seconds of each run were excluded.
Data processing and analysis were performed using SPM8, freesurfer, and custom
software. Data were motion corrected, realigned, normalized onto a common brain
space (rigid rotation and translation, MNI template), spatially smoothed using a
Gaussian filter (FWHM 5mm) and subjected to a high-pass filter (128 Hz).
Motion and Artifact Analysis
To estimate motion and data quality, we used three measures for each participant:
mean translation was defined as the average absolute translation for each TR across
the x, y, and z plane; mean rotation was defined as the average absolute rotation per
TR across yaw, pitch, and roll; and number of outliers per run was defined as the
average number of time points per run in which (a) TR-to-TR head movement exceeded
2mm of translation, or (b) global mean signal deviated by more than three standard
deviations of the mean. All outlier time points were excluded from both magnitude and
pattern analyses (using an impulse regressor of one TR). Runs missing more than 1/3
of the total time points due to artifact or motion were dropped, and participants with
fewer than 4 (of 5) remaining runs were dropped. Using these criteria, two participants
from Experiment 1 were excluded from all analyses.
fMRI Analysis
Modeling
All fMRI data were modeled using a boxcar regressor, convolved with a standard
hemodynamic response function (HRF). A general linear model was used to analyze the
BOLD data from each subject, as a function of condition, using a slow event related
design. The model included nuisance covariates for run effects, global mean signal, and
an intercept term, and impulse regressors to remove motion-artifact outliers. The
localizer was modeled as a 14s boxcar (the full length of the story and question period).
For the main experiment, we modeled the key sections of the stories separately: the
task was modeled using a single regressor for the first 16s of background (one
regressor for all 5 conditions), a single regressor for the conclusion and response time
(one regressor across all 5 conditions), and 5 condition regressors, modeling the 6-14
seconds of manipulated story (the “evidence” portion of the epistemic conditions).
Whole-brain random effects (univariate)
A second-level random effects analysis was performed on the contrast images
generated for each individual in the localizer task, by performing Monte Carlo
permutation tests using SnPM5 (Hayasaka & Nichols, 2004; http://www.sph.umich.edu/
103 of 179
ni-stat/SnPM/; T. Nichols, 2002), with a false positive rate controlled at p < 0.05, with
5000 permutations and a variance smoothing of 3mm.
Defining Individual subject ROIs
The whole-brain random effects analysis revealed eight main regions that showed
greater activation for false belief stories compared to false photograph stories (corrected
p < 0.05, pseudo-T=5.44): right and left temporo-parietal junction (RTPJ, LTPJ), right
middle superior temporal sulcus (RMSTS), right and left anterior superior temporal
sulcus (RASTS, LASTS), medial precuneus (PC), dorsal medial prefrontal cortex
(DMPFC), middle medial prefrontal cortex (MMPFC), ventral medial prefrontal cortex
(VMPFC).
Region
MNI
Size (Voxels)
Peak T
# Found
RTPJ
[54,-54,19]
277±67
11.01±2.44
17/17
LTPJ
[-47,-61,23]
251±81
9.69±2.05
17/17
RMSTS
[58,-25,-8]
158±76
8.19±1.71
17/17
RASTS
[56,-2,-18]
129±52
7.92±1.69
17/17
LASTS
[-54,1,-22]
123±88
6.56±1.74
17/17
PC
[1,-56,36]
300±63
9.60±2.00
17/17
DMPFC
[1,55,26]
145±102
7.57±2.29
15/17
MMPFC
[5,58,14]
157±108
7.87±2.33
16/17
VMPFC
[0,54,-14]
125±89
6.71±2.27
14/17
Table 4.2: Individual subject ROIs. Individual subject ROIs are defined as all voxels within a 9mm
sphere around the peak voxel (T of mental-physical from the localizer) identified in each region. MNI:
average coordinates of the peak voxels, across participants, reported in Montreal Neurological Institute
(MNI) space. Size: average number ± standard deviation of voxels included in individual ROIs. Peak:
average T value ± standard deviation of peak voxel for mental>physical. # Found: number of participants
in which an ROI was identified.
Based on the results of the whole-brain analysis, these functional regions of interest
(ROIs) were defined for each individual. For each participant, we masked the individual
T-map (one sample T-test of mental>physical) with hypothesis spaces defined from
random effects analysis on independent subjects (Dufour et al., 2013). In each
hypothesis space, we found the peak voxel in the contrast image for each region. ROIs
were then defined as all voxels within a 9mm radius of the peak voxel that passed
threshold in the contrast image (Belief > Photo, p<0.001 uncorrected, k>10 contiguous
voxels). Identifying the set of brain regions that are considered a core part of the theory
of mind network, these results replicate a number of studies using a similar functional
localizer task (Fletcher, Happé, Frith, & Baker, 1995; Gallagher et al., 2000; Gobbini,
Koralek, Bryan, KJ, & Haxby, 2007; Goel, Grafman, Sadato, & Hallett, 1995; Perner,
104 of 179
Aichhorn, Kronbichler, Staffen, & Ladurner, 2006; Saxe & Kanwisher, 2003; Van
Overwalle & Van Overwalle, 2009). These nine ROIs were identified in most individual
subjects (Table 4.2).
Within-ROI Multivariate Analysis (MVPA)
To ask whether any theory of mind ROIs contained information about our condition
manipulations, we conducted within-ROI multivoxel pattern analysis (MVPA) in
individual subject ROIs. The data from each individual ROI were classified using a
cross-validated linear support vector machine (SVM), which plausibly models the
readout of neural populations (Bialek, Rieke, Van Steveninck, & Warland, 1991; Butts et
al., 2007; DiCarlo & Cox, 2007; Hung, Kreiman, Poggio, & DiCarlo, 2005; Naselaris,
Kay, Nishimoto, & Gallant, 2011; Rolls & Treves, 2011; Seung & Sompolinsky, 1993;
Shamir & Sompolinsky, 2006), and performs as well as most non-linear classifiers (Bray,
Chang, & Hoeft, 2009; Cox & Savoy, 2003).
An individual ROI (Figure 4.2 A) had to have at least 20 voxels to be included in the
analysis. In each voxel, the data were high-pass filtered (within run; 128 Hz) and linearly
detrended (across runs). For each trial, we calculated the average BOLD signal in each
voxel in the ROI, measured during the “evidence” portion of the story (starting at 17 s
into the story, offset by 4s to account for hemodynamic lag, ranging from 4-16 s in
length). To normalize voxel response, BOLD responses in each voxel were z-scored
relative to the mean signal in that voxel across trials. This procedure resulted in a neural
“pattern” for each trial: a vector of the z-scored BOLD responses in each voxel inside
the ROI (Figure 4.2 B).
To reduce the dimensionality of the number of features (voxels) relative to the number of
items, feature selection was used to find voxels most likely to contain task-related signal
(De Martino et al., 2008; T. M. Mitchell et al., 2004; Pereira & Mitchell, 2009). Using data
from all runs, a task vs rest ANOVA identified the 100 voxels in each ROI in which the
response to all conditions differed most reliably from rest (ranked by F-statistic; T. M.
Mitchell et al., 2004). Because all conditions are modeled together, this selection
criterion did not bias the outcome of classification on condition pairs (“peeking”) and
allowed us to use the same voxels across cross-validation folds.
In this analysis, each item is treated as a single point in high-dimensional voxel space
(n=100); linear SVMs find a hyperplane through this space which maximizes the
distance between labeled classes (the margin), while minimizing the number of mislabeled data points. We conducted all analyses with a “soft” fixed regularization
parameter (C=1), allowing for a wider margin, at the cost of more mislabeled data in the
training set. After the model is built, new data can be classified based on the neural
pattern they elicit. Thus, successful classification requires that items generate neural
patterns that are reliably similar within a condition, and relatively different between
conditions.
The SVM algorithm was implemented with libSVM (http://www.csie.ntu.edu.tw/~cjlin/
libsvm/; C.-C. Chang & Lin, 2011). The data were partitioned into cross-validation folds
by run. A classification model was built iteratively on all runs but one (using labeled
data), and then used to predict the condition labels of items in the remaining run.
105 of 179
(Figure 4.2 C). Accurate predictions matched the actual stimulus label, and inaccurate
ones did not.
To determine subject-wise
c l a s s i fi c a t i o n a c c u r a c y,
accuracy was averaged across
items within a run, and then
averaged across runs, to yield
a single classification accuracy
for each participant. A region
could reliably distinguish
between two conditions if
subject-wise classification
accuracies were significantly
above chance (.50; onesample, one-tailed t-test12).
Figure 4.2 (A) RTPJ (B) Training and testing data. Matrix
shows mean activation in each voxel (rows) to each item
(columns). A classification model is built from the labeled
training data. The model finds a linear hyperplane (solid line)
through voxel space that best separates the two conditions
(light and dark dots) based on their neural activation (dashed
lines). The model is then applied to the left out testing data
(circled dots). Test items are labeled based on which side of
the hyperplane they fall; accuracy is computed as the
percentage of correctly labeled test items. (C) A two-voxel
model with 100% classification accuracy in the testing data; (D)
A two-voxel model with 75% classification accuracy.
12
Item-Wise classification:
Te s t i n g f o r c o n t i n u o u s
representations
In addition to looking at
subject-wise classification
accuracies, we also calculated
item-wise classification for
each condition. For each item,
classification of quality is
defined as the proportion of
times an item was classified as
“good evidence”, across
participants; modality
classification is defined as the
proportion of time an item was
classified as visual evidence.
Based on the subject-wise
results, we asked whether
ROIs that show a binary
distinction (e.g., successful
classification of good versus
bad evidence quality across
subjects) also represent graded
Parametric tests are not always appropriate for testing classification accuracy (Stelzer, Chen, & Turner,
2013). Here however, all classification data were independent (from separate subjects, rather than from
folds on overlapping data), and classification accuracies were normally distributed around the mean for
most ROIs. Where there was significant non-normality (using Lilliefor’s test for normality) in an ROI, we
report non-parametric statistics.
106 of 179
information about the distinction (e.g., very good evidence versus somewhat ambiguous
evidence across items).
We asked if item-wise classification scores (e.g., proportion of times an item was
classified as “good”) were predicted by item-wise behavioral ratings (1-7 scale of “how
good is the evidence”?). Because of the high collinearity between behavioral ratings and
condition labels, we used regularized linear regression (Hastie, Tibshirani, & Friedman,
2011; Hoerl & Kennard, 1970; ridge regularization; Krämer, Schäfer, & Boulesteix,
2009). Like multiple linear regression, ridge regression learns a beta weight for each
regressor, and uses the weighted sum of regressors to predict an outcome. However,
ridge regression simultaneously minimizes the sum of the squared prediction error and
the sum of the squared beta — effectively limiting the size of the coefficients, which are
often inflated by collinear regressors. In each model, the regularization parameter λ
determines the bias towards solutions with smaller beta values. As λ goes to 0, the
algorithm becomes identical to multiple linear regression (OLS). To avoid over-fitting λ,
for each model, we used cross-validation, iterating across 8 folds. Items were evenly
and randomly distributed into folds. In each iteration, items from one fold were set aside
for testing. In the remaining items, leave-one-out cross validation was used to build a
series of regularized linear models. Ridge regression was performed on the left-in
training items, using 1000 logarithmically spaced values of λ, ranging from from 1*10-8
to 30, implemented in R 3.1.1 (http://www.R-project.org R Core Team, 2013) using
packages parcor (Krämer et al., 2009) and ridge (Cule, 2012; Cule & De Iorio, 2013).
For each of these models, the predicted value for the withheld training item was
calculated. For each value of λ, the total residual was calculated as the sum of the
square of the predicted value minus the actual value for the withheld training items,
averaged across folds. The value of λ that minimized the overall error in the withintraining cross-validation was used to model the testing data. To determine how well the
model fit overall, betas for each regressor (condition labels and behavior) were
averaged across folds.
"
107 of 179
Results
fMRI task: Behavioral Results
After each story, participants answered whether the main character in the story felt
happy or sad. There were no differences in accuracy or response time between Seeing
Good-Evidence stories and Seeing Bad-Evidence stories (Accuracy (mean proportion
correct ± standard error): SG: 0.94±0.02; SB: 0.92±0.02; t(34.6)=0.55, p=0.59; RT
(mean time in seconds ± standard error): SG: 1.38±0.09; SB: 1.52±0.11; t(36.9)=-1,
p=0.32), between hearing first-person and hearsay (Accuracy: HF: 0.92±0.02; HS:
0.96±0.02; t(35.2)=-1.46, p=0.15; RT: HF: 1.39±0.08; HS: 1.37±0.09; t(37.6)=0.22,
p=0.83), or between seeing and hearing (Accuracy: seeing: 0.93±0.02; hearing:
0.94±0.02; t(38)=-0.45, p=0.65, d=-0.14; RT: seeing: 1.45±0.09; hearing: 1.38±0.09;
t(37.6)=0.54, p=0.59).
fMRI task: Classification Results
The key question we asked in Experiment 1 was whether any of the classic “ToM brain
regions” contains distinct spatial patterns of response to different features of others’
beliefs. Confirming that these regions play a role in ToM, 7/9 regions of interest showed
reliably distinct patterns of neural responses to stories describing the protagonist’s belief
versus stories describing physical events: RTPJ (accuracy=0.67(0.03), t(16)=5.7,
p<0.001), LTPJ (accuracy=0.61(0.04), t(16)=2.7, p=0.0074), RMSTS (accuracy =
0.62(0.03), t(16)=3.6, p=0.001), RASTS (accuracy = 0.57(0.03), t(16)=2.4, p=0.015), PC
(accuracy=0.62(0.03), t(16)=4.1, p<0.001), DMPFC (accuracy=0.59(0.03), t(14)=3.3,
p=0.0029), and MMPFC (accuracy=0.57(0.04), t(15)=1.8, p=0.049), with a marginal
effect in LASTS (accuracy=0.55(0.03), t(14)=1.7,p=0.06; Figure 4.3 A).
We next tested whether these regions distinguish between different features of beliefs,
including three epistemic distinctions (i.e. those that are relevant to the source of
others’ knowledge), and one salient but non-epistemic distinction (the valence of the
protagonist’s resulting emotion).
Quality
First, we tested whether any region represents others’ beliefs in terms of the quality of
evidence. Overall, the pattern of neural responses could successfully classify stories
describing good versus bad visual evidence in two regions: RTPJ (accuracy=0.56(0.03),
t(16)=2.4, p<0.014) and RMSTS (accuracy=0.58(0.04), t(16)=2.2, p=0.023, Figure 4.3
B), but in no other region (all p>0.1).
Independent behavioral ratings of the quality of the evidence in each story predicted
how likely that story was to be classified as depicting good evidence in both RTPJ
(r(78)=.29, p=0.009, Figure 4.3 C), and RMSTS (r(78)=.28, p=0.01, Figure 4.3 C). In
both regions, the behavioral ratings contributed significant information to neural
classification accuracies even after accounting for the binary condition labels (RTPJ:
Condition: beta=0.16±.1, t=1.7, p=0.11; Behavioral rating: beta=0.22±.1, t=2.3, p=0.038;
RMSTS: Condition: beta=0.2±.08, t=2.2, p=0.035; Behavioral rating: beta=0.17±.08,
t=2.1, p=0.048).
108 of 179
These results suggest that RTPJ and RMSTS not only encode the distinction between
good and bad evidence overall, but also encode continuous differences in evidence
quality. Interestingly, RMSTS and RTPJ seem to encode different aspects of the quality
of evidence depicted in the stories: there was no correlation between the classification
accuracies in RTPJ and RMSTS item-wise (r(78)=.18, p=0.12) or subject-wise
(r(15)=-0.19, p=0.48), and the classification accuracy of RTPJ and RMSTS made
independent contributions to the behavioral ratings (RTPJ: beta=1.0±.4, t=2.3, p=.02;
RMSTS: beta=1.1±.5, t=2.2, p=.03).
Modality
Second, we tested whether any region represents others’ beliefs in terms of the
modality of evidence. Replicating a prior study (Chapter 3, Koster-Hale, Bedny, & Saxe,
2014), we found distinct patterns of response to beliefs based on visual versus aural
evidence in RTPJ (accuracy=0.55(0.03), t(16)=2.1, p=0.027) and LTPJ
(accuracy=0.59(0.03), t(16)=2.8, p=0.007, Figure 4.3 B). The RTPJ and LTPJ appear to
contain very similar representations of the modality of evidence: classification scores for
modality in RTPJ and LTPJ were correlated across items (r(157)=0.21, p=0.009). Also,
across participants, individual differences in classification accuracies for modality were
correlated between RTPJ and LTPJ (r(15)=0.51, p=0.04). There were marginal trends
towards successful classification by modality in PC (accuracy=0.54(0.03), t(16)=1.4,
p=0.093) and in MMPFC (accuracy=0.53(0.02), t(15)=1.5, p=0.072).
Quality vs Modality
The RTPJ thus showed significant classification accuracy for both quality and modality
of evidence. These two dimensions are often correlated in real-world situations (e.g. the
expression “seeing is believing”), although they were deliberately made orthogonal in
our stimuli, so one possibility is that the RTPJ actually encodes only one of these two
dimensions, and the other dimension is distinguished by proxy. We tested this
hypothesis in two ways. First, we asked whether an item that was rated as depicting
better evidence was more likely to be classified as “visual” based on the pattern of
response in the RTPJ. We found no such correlation, either in all 160 items, or within
the 80 visual items alone (all items: r(158)=0.04, p=0.6; visual items: r(78)=-0.01,
p=0.9). Behavioral ratings of evidence quality were better predictors of neural
classification by evidence quality than by evidence modality (difference of correlations,
z(78)=1.93, p=0.05). Second, we asked whether there were shared individual
differences, across participants, in classification accuracies for evidence quality and
evidence modality in the RTPJ. We found no relationship between the classification
accuracy in a participant’s RTPJ for modality versus quality (r(15)=-.22, p=0.4).
Together, these results suggest that RTPJ contains independently significant and
orthogonal representations of the quality and modality of the source of others’ beliefs.
Directness
Third, we tested whether any region distinguishes between others’ beliefs based on
first-person (auditory) evidence versus hearsay (another person’s report), independent
109 of 179
Figure 4.3 (A) Individual ROIs (B) Classification accuracy in Experiment 1 for Mental: mental vs.
physical stories, Quality: good vs. bad quality evidence (SG, SB), Modality: visual vs auditory
modality (S,H), Directness: first person vs hearsay (HF, HS), and valence: positive vs negative
emotion, (C) Classification accuracy in the replication study for Quality: good vs. bad quality
evidence. (D). RTPJ, RSTS, VMPFC: Correlation between item-wise classification scores and
behavioral ratings. LTPJ: density plot of classification accuracies by condition. In RTPJ and
RMSTS there were significant correlations between quality classification scores (how many times
an item was scored as “good”, across participants) and behavioral judgments (how strong is the
character’s evidence?). In VMPFC, there was a significant correlation between valence
classification score (how many times an item was scores as “positive”) and behavioral judgements
(how happy does the character feel?). Modality is not a continuous feature and thus not compared
to behavior; rather we observe striking similarity in the classification accuracy in LTPJ between
seeing conditions (red: SG and SB) and hearing conditions (purple: HF HS). * p < 0.05
110 of 179
of quality and modality. However, we could not classify stories into these categories in
any region (all p>0.14).
Valence
Finally, we tested whether any ToM region distinguishes stories in which the protagonist
feels happy versus sad as a result of their conclusion. The pattern of neural response in
one region, the VMPFC, could reliably classify stories resulting in positive versus
negative emotions (accuracy=0.57(0.04), t(12)=1.9, p=0.038, Figure 4.3 B).
Independent behavioral ratings of the valence of the protagonist’s emotion in each story
predicted how likely that story was be to be classified as positive valence by the VMPFC
(r(158)=0.21, p=0.009, Figure 4.3 C). The behavioral ratings contributed significant
information to neural classification accuracies even after accounting for the binary
condition labels (Condition: beta=0.13±.07, t=2.2, p=0.047; Behavioral rating:
beta=0.15±.07, t=2.2, p=0.02).
Generalization to other manipulations of evidence quality
A key novel result of Experiment 1 is the evidence of a representation of the quality of
others’ evidence in RTPJ and RMSTS. To test the replicability of this result, and its
robustness to variation in the stimulus and task, we reanalyzed published data that
contained a similar manipulation.
Young et al. (2010) asked participants to make moral judgments of actions (e.g. “Grace
puts the powder in her friend’s coffee”) based on justified or unjustified beliefs. Belief
justification was manipulated by describing the belief as based on good evidence (e.g.
“Grace thinks the white powder is sugar, because the container is labeled ‘sugar’”) or
based on insufficient evidence (e.g. Grace thinks the white powder is sugar, though
there's no label on the container.”). To look for neural representations of belief
justification, Young et al. (2010) analyzed the average magnitude of response in ToM
brain regions, and found no difference in the magnitude of response to justified versus
unjustified beliefs in any region. However, they did not conduct any pattern analyses. In
the current re-analysis, we tested whether the pattern of response in any ToM region
(particularly the RTPJ and RMSTS) distinguished between beliefs based on good
versus insufficient evidence.
Stimuli
In this study, the “bad evidence” consisted not of obscured and ambiguous evidence,
but of missing or misleading evidence (Figure 4.4). In some stories, the protagonist
came to a conclusion without any evidence at all (believing a tree house is safe, despite
having never been inside, or believing that a watch is waterproof without seeing any
markings on it). In some stories, the protagonist is inattentive to key details (failing to
notice that someone next to them is choking); in others, the protagonist seems to
actively ignore good evidence (failing to read a note from a pet-owner specifying what
their kittens can eat, or failing to slow down to check whether a runner stopped on the
side of the road needs help). Finally, in some stories, the protagonist draws a
111 of 179
conclusion that seems to go directly against the available evidence (believing that a
sleeping roommate has heard a fire alarm even though he didn’t move, or believing
lunch meat is safe to eat despite its bad smell.) The manipulations of evidence quality in
this experiment are very different than Experiment 1, which manipulated the
protagonist’s perceptual access and the extent to which the evidence unambiguously
supported the conclusion. Thus, we can use these data to ask not only whether the
neural representations are found in a new set of participants, with new stimuli and a
new task, but also whether they extend to conceptually different manipulations of
evidence quality.
Good*Reason
Bad*Reason
Background
(6s)
Ray'and'his'girlfriend'are'on'a'weekend'hike.'They'come'across'a'narrow'bridge'that'spans'a'
rocky'creek.'
Evidence
(4s)
Ray'thinks'that'the'bridge'is'safe'to'walk'across Ray'thinks'that'the'bridge'is'safe'to'walk'across
because-it-isthough-it’s-not
marked'with'the'same'nadonal'park'symbol'as' marked'with'the'same'nadonal'park'symbol'as'
all'the'other'footbridges.
all'the'other'footbridges
AcHon*
and*
Outcome
(6s)
Ray'encourages'his'girlfriend'to'walk'across'the'bridge.
The'bridge'is'actually'very'unstable.'It'is'a'makeshie'bridge'put'together'by'other'hikers'without'
much'care.'Unfortunately,'it'collapses,'and'Ray's'girlfriend'falls'and'shagers'her'ankle.'
Task*(4s)
How*morally*blameworthy*is*Ray*for*encouraging*
his*girlfriend*to*walk*across*the*bridge?
Figure 4.4: Example stimuli from replication.
Methods and Analyses
In this study, 20 participants (12 women; mean age ± SD, 21.25 years ± 2.6) read 54
stories in the scanner. Each story was broken into three sections: (1) the background,
which set the stage (identical across conditions; 6 seconds), (2) the agent’s belief
(identical across conditions) and the justification for the belief (bad, good or
unspecified; 4 s), and (3) the conclusion, which detailed the agent’s action and the final
outcome (bad or neutral; 6 s; Figure 4.4). Only stories describing good or bad evidence
are analyzed here. Also, we collapse across outcome: 1/3 of the stories ended with a
bad outcome (the protagonist’s action led to physical injury or death of another
character), and 2/3 of the stories ended with a neutral outcome (no harm to anyone).
Each story appeared once in each condition, forming a matched and counterbalanced
design. Individual participants saw each story in only one of the 9 conditions; every
story occurred in all conditions, across participants. Stories were presented in a
pseudorandom order; conditions were counterbalanced across runs and participants,
and 14 s rest blocks were interleaved between stories. After each story, participants
responded to the question “How morally blameworthy is [the agent] for [performing the
action]?” on a 4-point scale (1-not at all, 4-very much), using a button press. Stories
were presented in white, 24-point font on a black background. Nine stories were
112 of 179
presented in each 5.1 minute run, with six runs total (30.6 minutes). Each participant
also completed the theory of mind localizer (four runs, 20 stimuli per condition).
Data were collected and analyzed using the same parameters as Experiment 1, though
with a 12 channel head coil. No participants were excluded due to motion. Individual
ROIs were picked using the data generated by the ToM localizer (Table 4.3). SVM
classification was performed in individual ROIs, on the data generated during the 4
second evidence section of each story (offset 4 s for hemodynamic lag).
Results
As in Experiment 1, the neural
pattern of response in the RTPJ
could successfully classify beliefs
based on good versus insufficient
evidence (accuracy=0.54(0.02),
t(17)=1.8, p<0.05, Figure 4.3 B:
quality 2). No other ROI showed this
d i s t i n c t i o n ( a l l c l a s s i fi c a t i o n
accuracies < 51%, p>0.3), including
the RMSTS (accuracy = 0.47(0.03),
t(15)=-0.99, p=0.83; Wilcoxon
signed-rank W = 6, p=0.70 13).
RMSTS carried reliably less
information about evidence quality
than RTPJ (Wilcoxon signed-rank W
= 86.5, p=0.04). Moreover, RMSTS
carried significantly less information
than it did in Experiment 1 (Wilcoxon
signed-rank W = 75.5, p-value =
0.03), suggesting that RMSTS
carries reliably less measurable
information in the replication than in
Experiment 1.
Exp 2
MNI
Size
Coordinate
(Voxels)
s
Peak T
# Found
RTPJ
[54,-51,19]
231±97
7.22±1.65
18/20
LTPJ
[-54,-58,25] 173±111
6.05±1.42
19/20
RMSTS
[57,-26,-7]
89±55
5.37±1.37
19/20
RASTS
[55,2,-22]
64±45
5.12±0.8
15/20
LASTS
[-54,3,-25]
40±36
4.5±0.74
15/20
PC
[1,-57,36]
233±80
6.48±1.36
20/20
DMPFC
[3,54,29]
93±86
5.06±1.1
15/20
MMPFC
[6,59,14]
77±70
5.24±0.92
13/20
VMPFC
[2,56,-8]
57±58
4.6±0.96
13/20
Table 4.3: Individual subject ROIs. Individual subject
ROIs are defined as all voxels within a 9mm sphere
around the peak voxel (T of mental-physical from the
localizer) identified in each region. MNI: average
coordinates of the peak voxels, across participants,
reported in Montreal Neurological Institute (MNI) space.
Size: average number ± standard deviation of voxels
included in individual ROIs. Peak: average T value ±
standard deviation of peak voxel for mental>physical. #
Found: number of participants in which an ROI was
identified.
13
Because the distribution of the RMSTS differed significantly from a normal distribution (Lilliefors test for
normality, K=0.26, criterion=0.21, p=0.004), we report results from the relevant parametric and nonparametric tests.
113 of 179
Discussion
Across two experiments, using the pattern of BOLD responses we could decode
whether another person’s belief was based on good versus bad evidence in the RTPJ.
In the first study, we also found evidence for a representation of the quality of others’
evidence in RMSTS. In the same paradigm, we replicated and extended two previously
observed findings: we could decode the perceptual modality of others' evidence (seeing
versus hearing) in the left and right TPJ, and positive versus negative emotional valence
in VMPFC. Overall, these results provide evidence for highly abstract representations of
others’ mental states, with distinct dimensions of mental states represented in different
regions within the “mentalizing network”.
Evidence Quality is explicitly represented in RTPJ and RMSTS
A key result of this paper is that the RTPJ and RMSTS contain a neural code which
tracks the degree to which someone’s evidence supports their beliefs. Using linear
classification, we found that the neural pattern generated by one set of story items
predicted the condition label in independent items, reliably distinguishing situations in
which someone comes to a conclusion based on evidence that strongly supports their
belief (good evidence) from situations in which someone comes to the same conclusion
based on unreliable and ambiguous evidence. Moreover, in both RTPJ and RMSTS, the
classification score of each item — the frequency with which it was labeled as “good”
evidence, based on the neural pattern it elicited — correlated with independent
behavioral ratings on the quality of the evidence. RTPJ and RMSTS each explained
significant and uncorrelated variance in the behavioral ratings, suggesting that the
regions encode different aspects of the stimuli. Together, these data indicate that the
neural populations in RTPJ and RMSTS contain explicit, and possibly complementary,
representations of a key epistemic feature of belief attribution: the quality of the
evidence that other people use to draw their conclusions.
Replicating Experiment 1, we found in a re-analysis of the data from Young et al. (2010)
that the neural pattern of response in the RTPJ reliably distinguished between beliefs
based on good versus insufficient evidence. The replication of this result in the second
experiment suggests that the effect in RTPJ generalizes across participants, stimuli, and
task. The converging results are particularly compelling because the two experiments
used different manipulations of evidence quality. While Experiment 1 manipulated
mainly perceptual aspects of evidence quality (e.g. distance, distraction, darkness,
clutter, brevity), the replication described relevant evidence as either present or absent,
or attended or ignored. Rather than being driven by superficial stimulus features or task
demands, the distinct neural patterns in the RTPJ thus likely reflect an underlying
distinction in the representation of the quality of others’ evidence for their beliefs.
By contrast, we found no evidence of a distinction between good and insufficient
evidence in the RMSTS in the replication. Although null results in MVPA must be
interpreted with caution (e.g., due to limits in spatial scale), this result is consistent with
the results in Experiment 1 showing that the RTPJ and RMSTS represent different
aspects of the quality of evidence. It is possible that the stimuli in the replication did not
114 of 179
include a manipulation of the key aspects of evidence quality represented in the
RMSTS. One interesting hypothesis is that RMSTS is sensitive to information about
perceptual ambiguity, while RTPJ tracks information about conceptual ambiguity. Or, the
RMSTS may be sensitive to features of the explicitly stated bad evidence or observers,
while the RTPJ more explicitly represents the link between the quality of the evidence
and the belief.
The modality of others’ evidence is explicitly represented in right and
left TPJ
Our second key finding is that the neural patterns in RTPJ and LTPJ code information
about the perceptual source of other people’s beliefs (did they see or hear the
evidence?), independent of quality or directness. These results replicate and clarify the
findings in Chapter 3 (Koster-Hale et al; 2014), who originally found that the neural code
in bilateral TPJ reliably distinguished the perceptual source (seeing versus hearing) of
other people’s beliefs. Here, we simultaneously manipulated modality, quality, and
directness. Finding converging results across experiments, despite differences in
stimuli, participants, and manipulation, suggests that the neural data are being driven by
the intended manipulation (modality), rather than other correlated but unintended
features of the stimuli. Interestingly (and unlike RTPJ and RMSTS), the information
represented in RTPJ and LTPJ seems very similar, across items and participants:
participants with high decoding accuracy in RTPJ had high decoding accuracy in LTPJ;
and items that had a high frequency of being classified as “seeing” in one region were
more likely to be frequently categorized as “seeing” in the other.
While RTPJ is implicated in the processing of others’ beliefs, relative to other types of
representations (such as maps or signs), LTPJ has been implicated in more general
meta-representation (tracking differences between two types of representation; for
example a person’s own memory of the geography, and a map of the area), and
possibly in processing misleading information and perspective taking in particular
(Aichhorn et al., 2009; Perner et al., 2006). Future research should test whether there is
a division of labor between these regions, and how they work together to represent
information about other people’s mental states. An intriguing possibility is that the LTPJ
may be implicated in general tracking of information sources, as part of encoding where
our information comes from, and keeping different pieces of information separate.
No region could decode directness: first-person experience versus
hearsay
The third epistemic dimension that we tested in our experiment was directness: whether
someone came to a conclusion based on their own, first person experience, or via
hearsay: information told to them by someone else. Directness plays an important role
in evaluating the experiences of others. Adults are more likely to trust another person if
they acquired their knowledge through first person experience than from hearsay (i.e.
other people’s report; Bull Kovera et al., 1992; Peter Miene et al., 1992; 1993; Robinson
et al., 1999) and children are less likely to be misled by adults reporting information they
learned via hearsay than via first person experience (Aydin & Ceci, 2009). Eyewitness
115 of 179
but not hearsay testimony is admitted as evidence in courts of law (Park, 1987).
Interestingly, however, we found no region that differentiated between direct evidence
and hearsay. This could be due to basic limits in the data (e.g. the representations are
not at a spatial scale measurable by our analysis). More interestingly, first person
experience versus hearsay may not be at opposite ends of a scale. If, for example, first
person experiences are represented as a heterogenous class (encoded via features
such as quality, modality, unexpectedness), there may not be a common neural
representation among items, and would not be detectable using pattern-based SVMs.
The valence of others’ emotions is represented in VMPFC
The final key result of this study is that the neural pattern in VMPFC codes information
about the valence of other people’s experiences — whether the protagonist felt positive
or negative at the end of the story. Moreover, each item’s valence classification score in
VMPFC correlated with independent behavioral ratings on the valence of each story.
These results suggest that VMPFC encodes continuous differences in emotional
valence.
Our results converge with and extend previous results on high-level emotion
representation in MPFC. Most relevantly, Peelen et al. (2010) found cross-modal
representations of emotion across facial expressions, body postures, and vocal
expressions in MPFC; and in the same region Skerry and Saxe (under review) found
evidence of a common neural code for others’ valence that generalizes across both
perceptual cues (facial expressions) and context-based inferences (short cartoons
without any perceptual cues to emotional valence).
Our results support the conclusion that VMPFC contains abstract representations of
valence. Our stories contain no perceptual cues (such as facial expressions), and, in
fact, very few emotion words, leaving the participants to infer valence from the context.
Moreover, to our knowledge, this is the first study to find distinct neural patterns for
others’ positive versus negative valence using text-based stimuli, and, more importantly,
the first study to find that these representations encode a continuous, rather than
categorical, representations of valence.
Neural representations of abstract features
Taken together, these results are particularly striking because they find evidence of
neural representations of highly abstract conceptual features. Information about other
people’s mind cannot be directly observed. Rather, we successfully reason about
others’ minds and actions by positing an internal causal structure of intentions, beliefs
and desires. The present results suggest that cortical regions implicated in social
cognition may represent features of this abstract causal model: the reasons why people
believe what they believe.
At least two of the features we tested (evidence quality and valence) are coded along a
continuous dimension. Previous work has found that, across participants, classification
accuracy can predict task performance (Kok, Jehee, & de Lange, 2012; Chapter 2,
Koster-Hale, Saxe, Dungan, & Young, 2013; Walther, Beck, & Fei-Fei, 2012) and clinical
symptoms (Coutanche, Thompson-Schill, & Schultz, 2011; Dosenbach et al., 2010),
116 of 179
suggesting that differences in neural patterns may reflect individual differences in
behavior. Here, we show that the probability of an item being categorized as one type of
stimuli is predicted by independent behavioral ratings, suggesting that the brain may
represent these stimuli along continuous, abstract feature dimensions (Kriegeskorte,
2008; Kriegeskorte & Kievit, 2013). A key next step will be to be to further characterize
the space that these dimensions are part of: what other features are represented, and
how are they coded?
Finally, these data suggest a division of labor across the brain. We find that information
about the source of others’ beliefs is represented in regions reliably implicated in social
cognition. Disassociations in functional and spatial profile in single individuals (Scholz,
Triantafyllou, Whitfield-Gabrieli, Brown, & Saxe, 2009) and meta-analyses (Schurz et al,
2014; Decety & Lamm, 2007; Kubit & Jack, 2013, Krall et al., 2014), functional
connectivity (Bzdok et al., 2012; Krall et al., 2014), and structural connectivity (Mars et
al., 2012) all suggest that both TPJ and MPFC contain neural populations that activate
selectively to mental state reasoning tasks, relative to a variety of other tasks. However,
a key open question is what kinds of computations these neural populations support.
MVPA has shown that RTPJ contains information about future social actions (Carter,
Bowling, Reeck, & Huettel, 2012), social groups versus individuals (Contreras,
Schirmer, Banaji, & Mitchell, 2013), and intent when causing harm (Chapter 2, KosterHale et al., 2013), and the reliability of others’ cooperation (Behrens, Hunt, Woolrich, &
Rushworth, 2008). In contrast, MPFC has been shown to track information about longterm personality traits (Hassabis et al., 2013) and emotional states (Kassam, Markey,
Cherkassky, Loewenstein, & Just, 2013; Peelen et al., 2010). The data here argue that
the RTPJ represents features of beliefs; and in particular the epistemic features of
beliefs, including the quality and modality of the evidence that gives rise to other
people’s beliefs.
This work adds to the growing body of literature that has found neural representations of
a variety of high-level cognitive domains. For example, person identity, a representation
that is highly abstract (we can determine the identity of someone by hearing their voice,
seeing their face, noticing their distinctive gait from afar, or seeing a corner of their
vintage purse), yet specific (each person we encounter is represented differently), is
encoded in anterior temporal lobe, generalizing across voice identification and facial
identification in a variety of environments, such as visual angle, lighting and partial
occlusion (Anzellotti et al., 2013). Information about personality traits, such as
extraversion, is tracked in MPFC (Hassabis et al., 2013). Value information, in both
monetary and social contexts, drives activity in VMPFC and ventral striatum (Behrens,
Hunt, & Rushworth, 2009; Izuma, Saito, & Sadato, 2008; Poore et al., 2012). Semantic
category is tracked in ATL (Correia et al., 2014). In combination, these types of results
suggest that the brain represents complex and abstract information in neural
populations of relatively circumscribed cortical patches — including in areas associated
with social cognition.
117 of 179
Why track epistemic features of beliefs?
Distinguishing reliable sources of information from unreliable ones informs key social
inferences: deciding what information we believe to be true, what actions to take, and
who we will trust. Critically, tracking the reliability of others’ beliefs plays an important
role in successfully learning from others. Much of our information comes from other
people; having information about the source of their beliefs can be key to knowing what
to keep and what to discard, while retaining the source of the evidence can allow us to
reevaluate our beliefs in the face of additional information — for example, learning that
an informant is nearly deaf allows us to retroactively update our confidence in the
gossip she shares (Robinson, Haigh, & Nurmsoo, 2008). Moreover, source information
allows the learner to probe for new knowledge by making inferences about what else
their informant is likely to know. For example a person who heard a glass smash, and a
person who saw the glass fall, both have good reason to believe that the glass is
broken, but only the person with visual access is likely to know that there is red wine on
the white carpet. Similarly, by age five years, children anticipate that a puppet who
looked at a toy will be more accurate about its color than a puppet who touched the toy,
or a puppet who is just guessing (O'neill & Chong, 2001; Pillow, Hill, Boyce, & Stein,
2000), and both adults and children selectively choose to learn new information from
reliable sources (Robinson & Whitcombe, 2003; Smith & Ellsworth, 1987; Toglia et al.,
1992; Whitcombe & Robinson, 2000), even when the speaker’s evidence contradicts
their own first person experience.
As well as informing our judgments about the reliability of the belief itself, source
knowledge can impact the kinds of inferences we draw about the long-term personality
traits of the belief-holder. Learning about the kinds of inferences that someone is willing
to draw, based on the evidence they observed, can be telling about their
trustworthiness, credibility, and belief system. When someone makes sensible
conclusions based on clear evidence, we are more likely to believe them in the future,
consider them trustworthy, and rely on them as a future source of evidence. In contrast,
when someone comes to a strong conclusion in the face of weak or unreliable evidence,
we are likely to conclude that they themselves are unreliable sources of information,
and possibly crazy, stupid, or untrustworthy (Clément, Koenig, & Harris, 2004; Hardwig,
1991; Koenig & Harris, 2005b; Lucas, Lewis, Pala, Wong, & Berridge, 2013; Olson,
2003). In fact, information about how someone acquired their beliefs can affect how
much we blame (and punish) them when things go wrong. When people make moral
judgements, causing harm by accident is typically judged to be much less blameworthy
than causing harm intentionally (e.g Cushman, 2008). If Grace puts a white powder in
her friend’s coffee, believing that the powder is sugar, she is less blameworthy than if
she knows the powder is poison (Young & Saxe, 2009). However, exoneration for the
accident depends substantially on why Grace thought the powder was sugar. If the
powder was in a jar labeled “sugar”, next to the coffee machine, then her belief was
reasonable and she is not to blame for the accident. If the white powder came from an
unlabeled jar in a chemical factory, then Grace is blamed much more for poisoning her
friend (Young et al., 2010).
118 of 179
An interesting open question is how sensitivity to source information develops. Tracking
the source of others’ beliefs develops along with other aspects of mental state
understanding well through the age of 10: children’s performance on theory of mind
tasks predicts their ability to report the sources of their own knowledge (Bright-Paul,
Jarrold, & Wright, 2008), their resistance to misleading information from uninformed
adults (Welch-Ross, Diecidue, & Miller, 1997), their ability to distinguish between
information from an informed source versus an informed source (Welch-Ross, 1999),
their ability to predict whether someone else will accept or reject conflicting information
(Robinson, 1994). Thus, finding representations of source in the adult brain provides a
key opportunity to probe the developmental trajectory of potentially late-developing
neural mechanisms of mental state inference.
Finally, an interesting open question is what role, if any, these same representations
play in representing one’s own experiences of good and bad evidence. Recalling how
one acquired one’s own information plays an important role in evaluating and justifying
our own beliefs, and deciding how readily they should be discarded (O'neill & Gopnik,
1991), and can improve memory fidelity (Dodson & Johnson, 1993; Lindsay & Johnson,
1989; Morton, 1994; Ross, Ceci, Dunning, & Toglia, 1994; Zaragoza & Lane, 1994).
Summary and Conclusion
In summary, using multi-voxel pattern analysis in two experiments, we find (i) that RTPJ
and RMSTS contain neural codes which track the degree to which someone’s evidence
supports their beliefs, reliably distinguishing between good, reliable evidence and
obscured and ambiguous evidence; (ii) that in both RTPJ and RMSTS, the frequency
that an item was labeled as “good” evidence was predicted by independent behavioral
ratings on the quality of the evidence, and that each brain region explained independent
variance in behavioral ratings; (iii), that RTPJ and LTPJ encode information about the
modality of other people’s evidence, (iv) that the neural pattern in VMPFC codes
information about the valence of other people’s experiences, and (v) that the frequency
of an item being classified as “positive” was predicted by independent behavioral ratings
of emotional valence.
Together, these data suggest that the brain regions implicated in social cognition contain
precise neural representations of abstract features, including epistemic and emotional
features of others’ beliefs, and that these features are represented in a continuous,
abstract feature space.
Acknowledgments
We thank the Athinoula A. Martinos Imaging Center at the McGovern Institute for Brain
Research at MIT, and the Saxelab, especially Amy Skerry and Stefano Anzellotti for
helpful discussion. We also gratefully acknowledge support of this project by an NSF
Graduate Research Fellowship (#0645960 to JKH) and an NSF CAREER award
(#095518), NIH (1R01 MH096914-01A1), and the Packard Foundation (to RS).
119 of 179
120 of 179
Chapter 5.
Theory of Mind: a neural prediction problem 14
Predictive coding posits that neural systems make forward-looking predictions about
incoming information. Neural signals contain information not about the currently
perceived stimulus, but about the difference between the observed and the predicted
stimulus. I propose to extend the predictive coding framework from high-level
sensory processing to the more abstract domain of theory of mind: that is, to
inferences about others’ goals, thoughts, and personalities. I review evidence that,
across brain regions, neural responses to depictions of human behavior, from
biological motion to trait descriptions, exhibit a key signature of predictive coding:
reduced activity to predictable stimuli. I discuss how future experiments could
distinguish predictive coding from alternative explanations of this response profile.
This framework may provide an important new window on the neural computations
underlying theory of mind.
14A
version of this chapter appeared as Koster-Hale, J., & Saxe, R. R. (2013). Theory of mind: a neural
prediction problem. Neuron, 79(5), 836–848. doi:10.1016/j.neuron.2013.08.020
121 of 179
Introduction
Social life depends on developing an understanding of other people’s behavior: why
they do the things they do, and what they are likely to do next. Critically, though, the
externally observable actions are just observable consequences of an unobservable,
internal causal structure: the person’s goals and intentions, beliefs and desires,
preferences and personality traits. Thus, a cornerstone of the human capacity for social
cognition is the ability to reason about these invisible causes: if a person checks her
watch, is she uncertain about the time or bored with the conversation? And is she
chronically rude or just unusually frazzled? The ability to reason about these questions
is sometimes called having a “theory of mind”.
Remarkably, theory of mind seems to depend on a distinct and reliable group of brain
regions, sometimes called the “mentalizing network” (e.g., Aichhorn et al., 2009; Saxe
and Kanwisher 2003), which includes regions in human superior temporal sulcus (STS),
temporo-parietal junction (TPJ), medial precuneus (PC), and medial prefrontal cortex
(MPFC). Indeed, the identity of these regions has been known since the very first
neuroimaging studies were conducted by 2000, based on four empirical studies, Frith
and Frith concluded that “studies in which volunteers have to make inferences about the
mental states of others activate a number of brain areas, most notable the medial
[pre]frontal cortex [(MPFC)] and temporo-parietal junction [(TPJ)]” (Frith and Frith 2000).
Since then, more than 400 studies of these regions have been published. However,
although there is widespread agreement on where to look for neural correlates of theory
of mind, much less is known about the neural representations and computations that
are implemented in these regions. The problem is exacerbated because these brain
regions, and functions, may be uniquely human (Saxe 2006; Santos et al., 2013).
Recent evidence suggests that there is no unique homologue of the TPJ or MPFC
(Rushworth et al., 2013; Mars et al., 2013), making it even harder to directly investigate
the neural responses in these regions.
In the current review, we import a theoretical framework, predictive coding, from other
areas of cognitive neuroscience and explore its application to theory of mind. There has
recently been increasing interest in the idea of predictive coding as a unifying
framework for understanding neural computations across many domains (e.g., Clark
2013). In this review, we adapt a version of the predictive coding framework that has
been developed for mid- and high-level vision. Like vision, theory of mind can be
understood as an inverse problem (Baker et al., 2011; Baker et al., 2009); the challenge
is to use the observable evidence (in this case human behaviors and states) to infer an
invisible causal structure that gave rise to the evidence (the goals, thoughts, and
personality of the individual; Seo and Lee 2012). Also like vision, theory of mind is a
complex cognitive process that depends on many different brain regions with likely
distinct computational roles (DiCarlo et al., 2012). We suggest that a predictive coding
framework can be used both to shed light on existing data about these brain regions,
and to suggest productive new lines of research.
First, we briefly review predictive coding, and sketch a model we believe can serve as
an integrative framework for the neuroscience of theory of mind. Second, we provide a
122 of 179
selective review of existing neuroimaging studies of theory of mind. Across different
stimuli and designs, with correspondingly different social information and predictive
contexts, we find a classic signature of a predictive error code: reduced neural response
to more predictable inputs. Third, we discuss how to distinguish predictive coding from
alternative explanations of this response profile, including differences in attention or
processing time. Based on recent neuroimaging experiments in visual neuroscience, we
suggest strategies for future experiments to test specific predictions of predictive
coding. Finally, we discuss the implications of predictive coding for our understanding of
the neural basis of theory of mind.
A predictive coding framework
The central idea of “predictive coding” is that (some) neural responses contain
information not about the value of a currently perceived stimulus, but about the
difference between the stimulus value and the expected value (Fiorillo et al., 2003;
Schultz et al., 1997; Schultz, 2010). This general idea is most familiar from studies of
“reward prediction error” in dopaminergic neurons in the striatum. Famously, these
neurons initially fire when the animal receives a valued reward, like a drop of juice, and
do not respond above baseline to neutral stimuli, such as aural tones. After the animal
has learned that a particular tone predicts the arrival of a drop of juice two seconds
later, the same neurons fire at the time of the tone. Tellingly, the firing rate of these
neurons no longer rises above baseline at the time the juice drop actually arrives.
Nevertheless, the neurons still respond to juice. If the tone that typically predicts a
single drop of juice is unexpectedly followed by two drops of juice, the neurons will
increase their firing; and if the tone is unexpectedly followed by no drops of juice, the
neurons decrease their firing rate below baseline (Fiorillo et al., 2003; Schultz et al.,
1997). These dopaminergic neurons exhibit the simplest and best known example of a
neural “error” code: the rate of firing corresponds to any currently “new” (i.e. previously
unpredictable) information about the value of coming rewards, not to the actual value of
any currently perceived stimulus (Bayer and Glimcher, 2005; Nakahara et al., 2004;
Tobler et al., 2005)
Predictive coding addresses a general challenge that an animal faces: developing an
accurate model of the expected value of all incoming inputs. Thus, predictive coding
models can be applied beyond the context of reward prediction to cortical processing
more generally. In fact, predictive coding was initially suggested as a model for visual
perception (Barlow, 1961; Gregory, 1980; Mumford, 1992), using a visual “error” code
that preferentially encodes unexpected visual information. The key benefit of such a
code, proponents suggest, is to increase neural efficiency, by devoting more neural
resources to new, unpredictable information.
By contrast to the single population of reward prediction error neurons, predictive coding
in the massively hierarchical structure of cortical processing poses a series of
challenges, however. If sensory neurons respond to errors, there must exist other
neurons to provide the prediction. Thus predictive coding models require at least two
classes of neurons: neurons that formulate predictions for sensory inputs (“predictor”
neurons, also called “representation” neurons; Summerfield et al., 2008; Clark 2013),
123 of 179
and neurons that respond to deviations
from the predictions (“error” neurons).
Because sensory input passes through
many hierarchically organized levels of
processing (DiCarlo et al., 2012;
F e l l e m a n a n d Va n E s s e n 1 9 9 1 ;
Logothetis and Sheinberg 1996;
Desimone et al., 1984; Maunsell and
Newsome 1987), a predictive model of
sensory processing requires an account
of the interactions between prediction
and error signals, both within a single
level and across levels.
To illustrate the idea, we provide our own
sketch of a hierarchical predictive coding
model. This proposal is a hybrid of
multiple approaches (Friston 2010; Clark
2013; Wacongne et al., 2012; de Wit et
al., 2010; Spratling 2010), seems to
capture the essential common ideas, and
is reasonably consistent with existing
data. The key structural idea is that
Figure 5.1. A sensory-coding based model of
social predictive coding. Predictor neurons (P) predictor neurons code expectations
code expectations about the identity of incoming about the identity of incoming sensory
input and pass down the prediction to lower level input and pass down the prediction to
predictor neurons (green arrows) and lower level both lower level predictor neurons and
error neurons (blue arrows). Error neurons (E) act lower level error neurons. Error neurons
as gated comparators, comparing sensory input act like gated comparators: they compare
from lower levels (red arrows) with the information
from predictor neurons (blue arrows). The sensory input from lower levels with the
difference between the predicted input and the information from predictor neurons. When
actual input is passed up higher level error neurons, the information that is being passed up
propagating up the processing hierarchy (red from lower levels matches the information
arrows). Error neurons also modulate the response carried by the predictor neurons, the error
of predictor neurons (purple arrows), likely both by neurons’ response to the input is
inhibiting the predictor neurons making incorrect
predictions, and enhancing predictor neurons reduced. This type of inhibition is the
making correct predictions. When the information classic signature of predictive coding,
that is being passed up from lower levels matches “explaining away” predictable input (Rao
the information carried by the predictor neurons, the and Ballard 1999). However, when
error neurons’ response to the input is reduced, predictor neurons at a higher level fail to
“explaining away” the predictable input (Rao and predict the input (or lack of input), there is
Ballard 1999).
a mismatch between the top-down
information from the predictor neurons
and the bottom-up information from lower levels, and error neurons respond robustly.
This error response propagates up the processing hierarchy. The consequence is a
sparse, efficient representation (mostly in predictor neurons) of predictable input, and a
124 of 179
robust, distributed response (mostly in error neurons) to unpredictable input, both
coordinated across multiple levels of the processing hierarchy (Figure 5.1).
Within a cortical region, population activity reflects a mixture of responses in the
predictor neurons (passing information about predicted inputs down the hierarchy) and
the error neurons (passing information about unpredicted inputs up the hierarchy). In
principle, predictive coding models need make no assumption about the distribution of
these two kinds of neurons within a population; in practice, aggregate population activity
is often dominated by error neurons (Friston 2009; Wacongne et al., 2012; Egner et al.,
2010; Keller et al., 2012; Meyer and Sauerland, 2009). The result is that the classic
signature of predictive coding, reduced activity to predictable stimuli, is typically
observed when averaging across large samples of neurons within a region (Meyer and
Olson 2011; Egner et al., 2010; de Gardelle et al., 2012). However, (as described in
more detail below) signatures of the predictor neurons can also be observed; for
example, the predictor neurons would likely show increased response when the input
matches their predictions (e.g., de Gardelle et al., 2012).
Following work in sensory processing (e.g., Wacongne et al., 2012), in our proposal
both error neurons and predictor neurons convey “representational” information, and
both are likely tuned to specific stimuli or stimulus features. Predictor neurons, present
at each level of the cortical hierarchy, do not code a “complete” representation of the
expected stimulus, but only some features or dimensions of the stimulus, at a relevant
level of processing. Each set of predictor neurons can explain only those particular
features or dimensions of the input, and correspondingly modulates the response in a
highly specific subset of error neurons. Error neurons are similarly distributed
throughout corticex, and respond to specific stimulus features (Meyer and Olson 2011;
den Ouden et al., 2012), rather than, for example, a single “error region” signaling the
overall amount of error or degree to which the observed stimulus is unpredicted (e.g.,
Hayden et al., 2011). Thus, for example, in the early visual cortex, predictor neurons
code information about the predicted orientation and contrast at a certain point in the
visual field, and error neurons signal mismatches between the observed orientation and
contrast and the predicted orientation and contrast. In IT cortex, predictor neurons code
information about object category; error neurons signal mismatches in predicted and
observed object category (den Ouden et al., 2012; Peelen and Kastner, 2011).
One consequence of this model is that, typically, the effects of predictions are limited to
relatively few levels of the processing hierarchy. To illustrate, expecting to see John walk
in the room would lead to predictions of biological motion, a body and a face (and
possibly to specific predictions within each of these domains), thus reducing error
responses in neural populations that respond at this level of abstraction. However, these
predictions often cannot be effectively or specifically translated into predictions at the
level of early visual receptive fields. Thus, while prediction signals may be passed down
the entire cortical hierarchy (Clark 2013; Rao and Ballard 1999), in many cases the
downstream transformation will make the signal too widespread to be informative. For
example, differential responses to predictable complex images have been observed in
monkey IT (Meyer and Olson 2011), and in human ventral temporal cortex (den Ouden
et al., 2010; Egner et al., 2010), without corresponding effects in lower visual areas.
Only when the environment supports specific, low-level predictions (on the scale of e.g.
125 of 179
orientation and contrast at specific points in retinotopy) should error signals be observed
at lower levels of processing (Alink et al., 2010; e.g., Murray et al., 2002; Weiner et al.,
2010).
When there is no relevant prediction available, error neurons act largely as “feature
detection” or “probabilistic belief accumulation” neurons (Drugowitsch and Pouget,
2012). This pattern highlights a key difference between predictive coding models
developed for sensory versus reward systems (den Ouden et al., 2012). Reward errors
are “signed”: the presence of an unexpected reward and the absence of an expected
reward are signaled by the same neurons changing their firing rates in opposite
directions. By contrast, sensory predictor neurons are likely “unsigned”: firing rates
increase in the presence of unexplained input.
Finally, our approach contrasts with other recent attempts to integrate social cognitive
neuroscience and predictive coding. Because predictive coding is most familiar from the
context of reward learning, there has been considerable interest in linking predictive
coding to social reward learning (Behrens et al., 2009; Jones et al., 2011; Fehr and
Camerer 2007). Social reward learning can mean either using social stimuli (e.g. smiling
faces) as rewards, or learning about rewards based on observation or consideration of
others’ experiences (Lin et al., 2012; Zhu et al., 2012; Zaki and Mitchell 2011; Poore et
al., 2012; Jones et al., 2011; Izuma et al., 2008; Chang and Sanfey, 2013; see Dunne
and O'Doherty 2013 for a review). Predictive coding may be an important mechanism
for motor control (i.e. anticipating, and explaining away, the consequences of one’s own
motor actions). Therefore some authors have linked motor predictions to social
predictions via the idea of “mirror neurons,” or shared motor representations for one’s
own and others’ actions (Brown and Brüne 2012; Kilner and Frith 2008; Patel et al.,
2012). The current proposal differs from both of these previous approaches by starting
with a hierarchical predictive coding framework developed for cortical visual processing
and by focusing on theory of mind, and specifically the attribution of internal states like
goals, beliefs, and personality traits.
This proposal is of course too general, and leaves many aspects of the model
unspecified (some of which we address below). Nevertheless, the basic features of
predictive coding described here provide an integrative framework for many findings in
the social cognitive neuroscience of theory of mind.
The sources of predictions
The social environment — the actions and reactions of other human beings — can be
predicted at a range of temporal scales, from milliseconds (where will she look when
the door slams?) to minutes (when she comes back, where will she search for her
glasses?) to months (will she provide trustworthy testimony in a court-case?). All of
these contexts afford predictions of a person's actions in terms of her internal states, but
the sources and timescales of the predictions are different. As we describe in the next
three sections, many experiments find that neural responses to predictable actions and
internal states are reduced, compared to unpredictable actions and states. This
common pattern can provide telling clues about the different types, and sources, of
126 of 179
predictions. We find that, while all regions show a higher response to unexpected
stimuli, what count as unexpected varies across regions and experiments, suggesting
that, at different levels of processing, neural error responses are sensitive to distinct
sources of social prediction.
To help clarify the sources of social prediction, we first review three sources of neural
predictions typically manipulated in visual cognitive neuroscience experiments. First,
given an assumption that the external world is relatively stable, neurons may predict
that sensory stimuli will remain similar over short timescales. Predictions based on very
recent sensory history can account for increased responses to stimuli that deviate from
very recent experience (Wacongne et al., 2012), and reduced responses to stimulus
repetition (Summerfield et al., 2008). Predictive coding may therefore offer an account
of widespread findings of repetition suppression in neural populations (Grill-Spector et
al., 2006). Predictive coding error is consistent with evidence that predictable repetitions
elicit more repetition suppression than unpredictable repetitions (Todorovic et al., 2011,
Todorovic and de Lange 2012).
Second, predictable sequences of sensory inputs can be created arbitrarily, through
training. For example, Meyer and Olson (2011) created associations between pairs of
images; for hundreds of training trials, image A was always presented before image B.
After training, the response in IT neurons to image B was significantly reduced when it
followed image A. This reduction was highly specific: the response remained high when
image B was presented alone, or following some other image, and there was no
reduction in the response to image A presented after image B. Other experiments show
that a tone can be used as a cue to predict the orientation of an upcoming grating (Kok
et al., 2012a), and either a colored frame or an auditory tone can predict whether an
upcoming image will be a face or a house (den Ouden et al., 2010; Egner et al., 2010).
The reliability of the cue can be stable over the experiment (Egner et al., 2010), or can
vary continuously across trials (den Ouden et al., 2010). In all cases, the magnitude of
neural responses tracks with the unpredictability of the stimulus, given the cue.
Perhaps most interesting, however, is the third source of predictions: an internal model
of the causal structure of the world that generated the observed input (Clark 2013;
Tenenbaum et al., 2011). For example, when two visual bars are presented in
alternating positions creating an illusion of motion, the visual system appears to
generate an internal model of a single object moving smoothly from one position to the
other across the intervening space. As a consequence, the addition of a third bar
presented at the right intervening space and time is treated as “predicted”, even though
that stimulus is otherwise unpredictable within the context of the experiment (Alink et al.,
2010).
In principle, all of three these sources of predictions can be applied to social prediction
and human actions. In practice, most of the experiments on theory of mind depend on
predictions based on prior expectations and an internal model of human behavior
(though we do find some evidence of predictions based on temporal proximity). Based
on the patterns of findings, we argue that these internal models must be quite abstract,
and include expectations that actions will be rational and efficient, and consistent with,
for example, the individual’s beliefs, personality traits, and social norms.
127 of 179
To reduce the complexity of this literature review, we focus here on three examples of
neural responses to actions at three conceptual levels: responses to biological motion
and goal-directed action in the superior temporal sulcus (STS), to other people’s beliefs
and desires in the temporo-parietal junction (TPJ) and to people’s stable personality
traits in the medial prefrontal cortex (MPFC; Figure 5.2). We find that, across all three
regions, with respect to the region’s preference and level of abstraction, expected
stimuli systematically elicit lower activation than unexpected stimuli.
Figure 5.2. Three brain regions involved in different aspects of theory of mind: examples of
individual regions of interest (ROIs). A region in posterior superior temporal sulcus (STS, green, peak
voxel [66, -36, 12]) involved in action perception (localized using biomotion relative to scrambled
biological motion, Pelphrey et al., 2003); a region in temporo-parietal junction (TPJ, blue, peak voxel
[62, -52, 18]) involved in thinking about beliefs and desires (localized using stories about mental states
relative to stories about physical events, Saxe and Kanwisher, 2003), and a region in medial prefrontal
cortex (MPFC, red, peak voxel [-2, 56, -4]) involved in thinking about people’s stable preferences and
personalities (localized using attribution of traits to self relative to attribution of traits to someone else,
Mitchell et al., 2006). All three ROIs localized using single subject data, p<0.001
Predicting goal directed action
The most immediate dimension of the social environment is the visibly observable
movements of other people’s bodies (e.g. grasping an item, running away) and faces
(e.g. gaze shifts, emotional expressions). Brain regions in the superior temporal sulcus
(STS) are implicated in many aspects of social action perception, showing robust
responses to face and body action in both humans (Jellema et al., 2000; Puce et al.,
1998; Bonda et al., 1996; Allison et al., 2002) and monkeys (Perrett et al., 1985; Jellema
and Perrett 2003b; Jellema and Perrett 2003a). Patients with lesions to the STS have
difficulty recognizing actions (Battelli et al., 2003; Pavlova et al., 2003), an effect that is
reproduced by creating reversible “lesions” in the STS through repetitive TMS
(Grossman et al., 2005). Consistent with a prediction error code, STS response to
observed actions is reduced when the observed action can be predicted, and enhanced
when the observed action is less predictable. These predictions appear to arise from a
variety of sources, ranging from experimental statistics, to constraints on biological
motion, to assumption about rational action, suggesting that rather than representing
low-level sensory-based statistics, this region represents (and makes predictions about)
coherent, rational actions.
128 of 179
First, like many sensory regions, the STS response is sensitive to the recent history of
the experiment and is reduced by repetition of a stimulus relevant to human action
perception. If two successive images of faces have the same gaze direction (i.e. both
gazing right) or the same facial expression (e.g. fearful), the STS response is reduced
compared both to a non-repeated presentation and to a repeated presentation of an
irrelevant stimulus, such as a house or object (Calder et al., 2007, Ishai et al., 2004, Furl
et al., 2007). Similarly, presenting the same action twice in row, from different viewing
angles, positions, sizes, and actors leads to reduced STS response relative to a
different action (Grossman et al., 2010; Kable and Chatterjee, 2006).
Human action can also be predicted based on internal models at many levels of
abstraction, from biomechanics to a principle of rational action. The most basic (and
most temporally fine-grained) predictions are constrained by the structure of bones and
joints and the forces exerted by muscles. Observers can thus predict the spatiotemporal
trajectory of human movements, especially for ballistic motions (Blake and Shiffrar
2007). Human movements that violate these biomechanical predictions (for example, a
finger bending sideways) elicit a higher response than more predictable movements in
the STS and related areas (e.g. Costantini 2005). Watching a human-like figure make
robot-like, mechanical movements elicits more activity than either a human-like figure
making human-like movements or a robot making mechanical movements (Saygin et
al., 2012).
Even when they do not violate biomechanical laws, human actions have a typical spatial
and temporal structure. Thus, if a person is walking rapidly across the room, we predict
that they will continue in the same trajectory, even if they are temporarily occluded. The
posterior STS responds more when the person reappears later than expected than
when the person emerges at the predicted time; when the person is replaced with a
passively gliding object, there is no effect of the time lag (Saxe et al., 2004).
In addition to intrinsic aspects of the action, observers expect others’ actions to be
temporally and spatially contingent on the structure of the environment. If a bright object
flashes near a woman’s head, she is very likely to immediately shift her gaze toward the
object. Seeing the woman immediately shift her gaze away from the bright object elicits
a higher response in the STS than the predicted gaze shift towards the object (Pelphrey
et al., 2003; Pelphrey and Vander Wyk 2011). This difference is reduced if the woman
first waits a few seconds before shifting her gaze, breaking the perception that that flash
caused the gaze shift. Similar effects are observed in infants as young as 9 months,
using EEG (Senju et al., 2006). In a more extreme mismatch between behavior and
environment, watching an agent twisting empty space next to a gear drives a stronger
STS response than the agent twisting the gear (Pelphrey et al., 2004).
Finally, the STS internal model of human behavior includes something like a principle of
rational action: the expectation that people will tend to choose the most efficient
available action to achieve their goal. The same action may therefore be predicted, or
unpredicted, depending on the individual’s goals and the environmental constraints
(Gergely and Csibra 2003). Correspondingly, the STS response is higher when the
same biomechanical action is unpredicted either because it is inefficient, or because it is
not a means to achieve the individual’s goal.
129 of 179
For example, action efficiency can be manipulated by having a person take a short or
long path to the same goal (Csibra and Gergely 2007), e.g. reaching for a ball efficiently,
arching her arm just enough to avoid a barrier, or inefficiently, arching her arm far above
the barrier. Across differences in barrier height and arm trajectory, activity in a region of
the MTG/STS is correlated with the perceived inefficiency of the action (Jastorff et al.,
2011). In a related experiment, observers watch someone performing an unusual action,
e.g. a girl pressing an elevator button with her knee. The context renders her action
more or less efficient: either her hands are empty, she is carrying a single book, or her
arms are completely occupied with a large stack of books. Activity in STS is highest
when the action appears least efficient, and lowest when the action appears most
efficient (Brass et al., 2007). The STS also responds more to failed actions (e.g., failing
to drop a ring onto a peg), an extreme form of inefficiency, than to successful ones
(getting the ring onto the peg, Shultz et al., 2011). Predictions for efficient action can
even be completely removed from the familiar biomechanics of human body parts: the
same inefficient action (going around a non-existent barrier) elicits stronger responses
in STS than the efficient version of the same action, when executed by a “worm” (a
string of moving dots, Deen and Saxe 2012).
In other experiments, the STS shows enhanced responses to actions that are
unpredicted given the individual’s specific goals, even if the action is not inherently
inefficient or irrational. For example, an individual who likes (and smiles at) a mug and
dislikes (and frowns at) a teddy bear can be predicted to reach for the mug and not the
bear. The goal-inconsistent action (reaching for the mug) elicits a higher response in the
STS (Vander Wyk et al., 2009). Similarly, when two people are cooperating on a joint
action, the STS shows increased responses when one person fails to follow the other’s
instructions: e.g. when asked to select one specific object (e.g., a red ball), the actor
takes the other object (e.g., the white ball; Shibata et al., 2011; see also Bortoletto et al.,
2011).
In sum, observers expect human movements to reflect “actions” which are sensitive to
the environment, and efficient means to achieve the individual’s goals. These
expectations can generate predictions for sequences of movements on the timescale of
seconds. All of these sources of predictions can modulate the neural response in the
STS, which is reduced when the stimulus fits the prediction.
Predicting beliefs and desires
Moving from the scale of seconds to the scale of minutes, the more general version of
the principle of rational action is that people will act efficiently to achieve their desires,
given their beliefs (Baker et al., 2011). Unlike specific motor intentions, beliefs and
desires last from minutes (e.g. the belief that your keys are in your purse) to years (e.g.
the desire to become a neurosurgeon). These beliefs and desires can be used to
predict aspects of a person’s actions, emotions, and other mental states, especially
when the person’s beliefs and desires differ from those of the observer (Wellman et al.,
2001; Wimmer and Perner 1983). Among other regions, a brain region posterior to the
superior temporal sulcus, in the temporo-parietal junction (TPJ), shows a robust
responses while thinking about an individual’s beliefs and desires (Saxe and Kanwisher
130 of 179
2003; Young and Saxe 2009a; Aichhorn et al., 2009; Perner et al., 2006). If the TPJ
includes a prediction error code, it should respond more strongly to beliefs and desires
that are unexpected, given the context. Indeed, there is evidence that the TPJ response
is reduced when a person’s beliefs and desires are predictable (though note that the
results reviewed in this section were generally not interpreted in terms of prediction
error coding by the original authors). In all of these experiments, the source of prediction
is not recent experimental history or trained associations, but rather a high level
generative model of human thoughts and behaviors.
One source of predictions about a person’s beliefs and desires is their actions (Patel et
al., 2012). Observers expect other people to be self-consistent and coherent (e.g.
Hamilton and Sherman 1996). This sensitivity to inconsistencies in belief and action is
reflected in the TPJ. For example, in one group of studies, participants read about an
act of violent harm or murder. Under the assumption that people usually act in
accordance with their beliefs (Malle 1999), the prediction is that the perpetrator intended
the harm; most assaults and murders are not accidental. Next, the participants read
about the perpetrator’s actual beliefs and desires. Responses in the right TPJ are higher
for “unpredicted” innocent or benevolent intentions that exculpate the harm (e.g. she
believed the poison was sugar; he only wanted to end the patient’s misery from an
incurable disease) compared to the “predicted” intention (to kill the person; Buckholtz et
al., 2008; Chapter 2, Koster-Hale et al., 2013; Yamada et al., 2012; Young and Saxe,
2009b).
Not all actions imply the corresponding intention, however: for example, violation of
social norms (e.g., spitting out a friend's cooking back on your plate) are more likely to
be committed accidentally than intentionally. Consistent with a prediction error code, the
TPJ response is higher for violations of norms performed intentionally ("because you
hated the food") versus unintentionally ("because you choked"; Berthoz et al., 2002).
In addition to these general principles, an individual’s beliefs and desires can
sometimes be predicted based on other information you have about his or her specific
group membership and social background. For example, Saxe and Wexler (2005)
introduce characters with different social backgrounds, ranging from the mundane (e.g.
New Jersey) to the exotic (e.g. a polyamorous cult). Participants then read about that
character’s beliefs and desires (e.g. a husband who believed it would be either fun or
awful if his wife had an affair). The response in right TPJ is reduced for the belief that
was predictable, given the character’s social background: the person from New Jersey
thinking his wife having an affair would be awful, and the person from the polyamorous
cult thinking his wife having an affair would be fun. Similarly, when reading about a
political partisan, political beliefs that are unexpected, given the individual’s affiliation
(e.g. a Republican wanting liberal Supreme Court judges) elicits a higher response in
right TPJ (Cloutier et al., 2011).
On the other hand, the general plausibility of a belief, in the absence of specific
background information about the individual, does not seem to be sufficient to generate
a prediction (or a prediction error) in the right TPJ. Without specific background
information about the believer, there is no difference in the right TPJ response to absurd
versus commonsense beliefs (e.g. “If the eggs are dropped on the table, Will thinks
131 of 179
they’ll bounce / break,” (Young et al., 2010), although the participants themselves rated
the absurd beliefs significantly more “unexpected.” A possible interpretation of these
results is that the internal model of the RTPJ predicts another person’s beliefs based on
expectations of a coherent individual mind, using information about that individual’s
specific actions and history, but not based on expectation that others will share one’s
own beliefs (see below for a contrast with MPFC) or common knowledge. However,
since these are null results, they should be interpreted with caution.
In sum, the response in the TPJ to other people’s beliefs and desires can be modulated
by how predictable those beliefs and desires are, relative to the current environment,
the individual’s actions, broader social norms, and the individual’s specific social
background.
Predicting preferences and personalities
At even longer timescales, successful prediction of the social environment depends on
building distinct models of each of the individual humans who compose one’s social
group. While some general rules, like the principle of rational action, apply to all people,
predicting a specific person’s action often depends on knowing the history and traits of
that individual. Brain regions on the medial surface of cortex, in both medial prefrontal
(MPFC) and medial parietal (PC) cortex, show robust responses while thinking about
people’s stable personalities and preferences (Mitchell et al., 2006; Schiller et al., 2009;
Cloutier et al., 2011). Consistent with a predictive error code, these responses are
reduced when new information about a person can be better predicted. Again these
predictions appear to be derived from relatively high level expectations that people’s
traits will be consistent across time and contexts, rather than from local experimental
statistics.
Prior knowledge of a person can be acquired through direct interaction. First person
experience of another person’s traits (e.g. trust-worthiness, reliability), can be
manipulated when participants play a series of simples ‘games’ with one or a few other
players. By gradually changing the other players’ behaviors, it is possible to create
parametric “prediction errors”. In one experiment, for example, the other player provided
“advice” to the participant; this advice shifted over the experiments, so that it was
reliable in some phases, and unreliable in others. The response in MPFC tracks with
trial-by-trial error in expectations about the informant’s reliability (Behrens et al., 2008).
Expectations about other people’s traits can also be based on verbal reports and
descriptions. For example, the initial behaviors of a (fictional) stranger can create an
impression of a certain kind of personality (e.g. ‘Tolvan gave her brother a compliment’).
The MPFC response is enhanced when later actions by the same person are
inconsistent with (i.e. unpredicted by) this trait (e.g. ‘Tolvan gave her sister a slap’)
compared to when they are predictable (e.g. ‘Tolvan gave her sister a hug‘; Ma et al.,
2012; Mende-Siedlecki et al., 2012).
When specific information about a person’s reputation or traits is unavailable, we may
predict others’ preferences by assuming that they will share our own preferences
(Krueger & Clement 1994; Ross, Greene, & House 1977). In one series of studies
(Tamir and Mitchell 2010), participants judged the likely preferences of strangers (e.g. is
132 of 179
this person likely to “fear speaking in public” or “enjoy winter sports”?) about whom they
had almost no background information. Under those circumstances, the response of the
MPFC was predicted by the discrepancy between the attributions to the target and the
participant’s own preference for the same items: the more another person was
perceived as different from the self, for a specific item, the larger the response in MPFC.
In all, human observers appear to formulate predictions for other people’s movements,
actions, beliefs, preferences and behaviors, based on relatively abstract internal models
of people’s bodies, minds, and personalities. These predictions are reflected in multiple
brain regions, including STS, TPJ, and MPFC, where responses to more predictable
inputs are reduced, and to less predictable inputs are enhanced.
Consistent with our general proposal for prediction error coding, reduced responses to
predicted stimuli in these experiments are typically restricted to relatively few brain
regions, and by implication, to relatively few levels of the processing hierarchy. Beliefs
or actions that are unpredicted, based on high level expectations, do not elicit enhanced
responses at every level of stimulus processing (e.g. early visual cortex, word form
areas, etc). Nor are prediction errors signaled by a single centralized domain general
“error detector”. Instead, relatively domain- and content-specific predictions appear to
influence just the error response at the relevant level of abstraction.
Beyond error
In sum, human thoughts and actions can be rendered unexpected in many ways, and
across many such variations a common pattern emerges: brain regions that respond to
these stimuli also show enhanced responses to “unexpected” inputs. This profile is the
classic signature of error neurons, and therefore consistent with a predictive coding
model of action understanding.
While consistent with predictive coding, however, these results provide only weak
evidence in favor of predictive coding. Increased responses to unexpected stimuli can
be explained by many different mechanisms, including increased ‘effort’ required,
increased attention, or longer evidence accumulation under uncertainty. The predictive
coding framework will therefore be most useful if it can make more specific predictions
and suggest new experiments.
Distinguishing prediction from attention
A salient alternative explanation for enhanced responses to unpredicted stimuli relies on
attention. Unexpected stimuli may garner more attention, and increased attention can
lead to more processing and higher activation (e.g., Bradley et al., 2003; Lane et al.,
1999). Similarly, increased processing effort or longer processing time can predict
higher activation (e.g. Cohen et al., 1997). Thus, higher activation to unexpected stimuli
could reflect greater attention or longer processing, rather than prediction coding errors.
However, relative to these accounts, predictive coding has a distinctive signature.
By hypothesis, predictions codes are more precise, more computationally efficient, and
less noisy than error codes (Friston 2005; Jehee and Ballard 2009; Rao and Ballard
1999; Spratling 2008). As a result, in a predictive coding model, better speed and
133 of 179
accuracy of perception are associated with reduced overall neural responses to
predicted stimuli (Kok et al., 2012a; den Ouden et al., 2008). By contrast, attention may
cause better speed and accuracy of performance by increasing overall neural
responses to attended stimuli (Feldman and Friston, 2010; Friston, 2010; Herrmann et
al., 2010; Hillyard et al., 1998; Kok et al., 2012b; Martinez-Trujillo and Treue, 2004;
Reynolds and Heeger, 2009; Treue and Martínez-Trujillo, 1999). That is, whereas
attention may increase gain in neural responses to the attended stimulus, predictions
improve perception by decreasing noise (or increasing sparseness) in neural responses
to the predicted stimulus.
If the neural responses described in the previous sections reflect prediction error,
reduced neural responses should be accompanied by improvements in behavioral
performance: people should make judgments more quickly, with less error, and with
more sensitivity on expected stimuli. Indeed, behavioral evidence suggests that
observers make faster and more accurate judgments about people who behave as
expected in social contexts. After watching two people engage in part of a cooperative
action or conversation, participants are faster and more accurate when both agents are
behaving as expected (e.g., responding aggressively or cooperatively, responding
communicatively or non-communicatively, or right away, instead of too early or late;
Manera et al., 2011; Neri et al., 2006; Graf et al., 2007). Important next questions will be
to look for these signatures in other aspects of social cognition, such as goal inference
or belief attribution.
Figure 5.3. Predicted stimuli elicit reduced activity with sharpened representations. Kok et al.
(2012a) report that an accurate prediction led to reduced response amplitude in early visual cortex, but
also simultaneously to an “improved” stimulus representation, as measured by multi-voxel pattern
analysis. Consistent with this suggestion, we find that in the right temporo-parietal junction (RTPJ; right
box) the response amplitude to a predicted belief was lower, but the spatial pattern associated with that
belief category was more reliable. Left: a sample stimulus. All stories described first a harmful action,
and then the agent’s belief. The “predicted” belief (solid arrow) was consistent with the action (i.e.
making the act an intentional harm). The “unpredicted” belief (dotted arrow) was inconsistent and
rendered the harm an accident. Middle: The amplitude of response in the TPJ was lower for the
intentional than accidental condition. Right: The spatial pattern of response in the TPJ was most robust
and reliable across trials for intentional harms, and somewhat less reliable for accidental harms. Data
from Chapter 2 and Koster-Hale et al. (2013).
134 of 179
An interesting extension of this idea is the proposal that the sparser prediction signal
should also be easier to decode from a neural population than the more distributed error
signal, within a single region and task (Kok et al., 2012a; Sapountzis, Schluppeck,
Bowtell, & Peirce, 2010). In an elegant study, Kok et al. (2012a) asked participants to
make fine perceptual discriminations between oriented gratings. They hypothesized that
when the orientation of the gratings was accurately predicted by a cue, the
representation of the grating would be largely in the sparser predictor neurons, whereas
when the orientation was not accurately predicted (i.e. on the relatively rare invalidly
cued trials), then the representation of the orientation would be largely in the more
distributed error neurons. Three predictions of their model were confirmed in the
responses of early visual cortex. First, the overall response to the gratings was lower
when the orientation was predicted than when it was unpredicted (the classic pattern of
“explaining away” the error signal). Second, behavioral discriminations on the gratings
were more accurate when the orientation was predicted than when it was unpredicted,
consistent with the hypothesis of a more efficient code. Third, and critically, the
orientation of the gratings could be more easily decoded from the spatial pattern of
neural responses in early visual cortex when the orientation was predicted than when it
was unpredicted, consistent with the hypothesis that the prediction signal is spatially
sparser than the error signal. Finally, Kok et al. (2012a) distinguished the effects of
prediction from effects of attention, by manipulating the participants' task. Directing
attention to the gratings’ orientation (versus contrast) improved decoding of orientation
in V1, but the effects of attending to orientation, and of seeing the unpredicted
orientation, were independent and additive.
A corresponding hypothesis should be easy to test with respect to the neural
representation of human behaviors, thoughts and personalities. The lower responses to
expected stimuli should be accompanied by better decoding of relevant stimulus
dimensions. Indeed, our own results from the TPJ are consistent with this hypothesis.
As described above, when reading about harmful actions (e.g. putting poison powder in
someone’s coffee), the TPJ response is higher to “unpredicted” innocent beliefs (e.g.
that the powder was sugar) than to “predicted” consistent beliefs (e.g. that the powder
was poison; Young and Saxe 2009b). We also found that using spatial pattern analysis
in the TPJ, we could decode the difference between innocent and guilty beliefs (Chapter
2, Koster-Hale et al., 2013). Based on Kok et al. (2012a), a further prediction is that the
decoding should be driven by a sparser and more efficient response to the predicted
category; and indeed, re-analysis of our data suggests that the guilty beliefs elicit a
more distinctive (i.e. more correlated across trials) spatial pattern than the “unpredicted”
innocent beliefs (Figure 5.3).
Interestingly, the benefits of an accurate prediction may be quite specific to the aspects
of the stimulus that are accurately predicted. As we suggest earlier, most predictions are
limited to a particular level of abstraction; given a high-level prediction, the probability of
lower-level features appearing will be too widely distributed to be informative. As a
result, accurate predictions may improve behavioral performance (and neural decoding)
at the representational level of the prediction (e.g. which object a person wanted) but fail
135 of 179
to improve, or even degrade, these measures for lower-level features (e.g. where in
space someone looked; He et al., 2012).
Finding the prediction signal
An important direction for future research will be to focus on signatures of the predictor
neurons, in addition to the error neurons. At least four different strategies may help to
identify prediction signals, and distinguish them from the often more dominant error
signals.
First, unlike error neurons, predictor neurons should show robust activity when the
stimulus fits prior predictions. Consistent with this suggestion, a recent study found that
although the majority of voxels in the fusiform face area (FFA, Kanwisher et al., 1997;
Kanwisher 2010) was suppressed for a repeated face, a subset of voxels reliably
showed the reverse pattern (de Gardelle et al., 2012), termed repetition enhancement
(see also Turk-Browne et al., 2006; Müller et al., 2013). Intriguingly, these two
populations of voxels also showed different patterns of functional connectivity. It will be
intriguing to test whether the STS, TPJ, PC, or MPFC similarly contain subsets of voxels
with enhanced responses to predicted actions or beliefs, and whether these voxels have
distinctive patterns of functional connectivity with other regions, especially because
unlike face processing, the direction of information flow among regions involved in
theory of mind is largely unknown.
Second, because both predictor neurons and error neurons may have preferred stimuli
(or stimulus features), it may be possible to identify the content of the prediction
independent from the response to the subsequent stimulus. For example, the response
of the FFA seems to increase when a face stimulus is predicted, as well as (and partially
independent from) when a face stimulus is observed (den Ouden et al., 2010; Egner et
al., 2010). Note though that neither of the existing studies could fully independently
identify the response to predicting a face, because in both cases, the probability of a
face was exactly reciprocal to the probability of the only other possible stimulus, a
house. By including a third category of stimulus, or a third possible cue, or by
independently varying the predictive value of the two cues, it should be possible to
independently measure category-specific responses to the prediction of a category,
versus the response to that category when observed.
Third, and relatedly, predictor neurons can signal the expectation of a stimulus that
never occurs. In some cases, the absence of an expected stimulus should generate
error activity as (den Ouden et al., 2010; Todorovic et al., 2011; Wacongne et al., 2011).
For example, the activity pattern in IT generated by the surprising absence of an object
contains information about the identity of the absent stimulus (Peelen and Kastner
2011). Unlike the “signed” (i.e. below baseline) error response in reward systems,
sensory neurons thus seem to show an increased response to an unexpectedly absent
stimulus (though note that there is some disagreement as to whether this activity is
driven only by the prediction signal before the stimulus is expected to appear, or by a
combination of the prediction signal with a subsequent error signal when the stimulus
fails to appear, e.g., den Ouden et al., 2010).
136 of 179
Fourth, the prediction and the error signals could be separable in time. Specifically,
under some circumstances prediction signals can begin before the stimulus, whereas
error signals are typically triggered by the stimulus itself (e.g., Hesselmann et al., 2010).
The low temporal resolution of fMRI may make it hard to test this hypothesis directly.
However one pattern of results is consistent with the idea that the STS contains
predictions of upcoming biological motion: still photographs of a person in mid-motion
(such as a discus thrower in the middle of throwing a disc) elicited more activity in the
STS than images that do not imply or predict motion (the same discus thrower at rest;
Kourtzi and Kanwisher 2000; Senior et al., 2000).
Fifth, error responses in a single region may be influenced by predictions from different
sources, and these different sources may be spatially separable. For example, FFA
shows repetition suppression for both repetition of one identical face image (plausibly a
very low-level prediction) and for repetition of a face across different sizes (requiring a
higher-level prediction). These error signals were related to different patterns of
functional connectivity between FFA and lower level regions (Ewbank et al., 2013). By
analogy, there may be different patterns of functional correlations related to different
sources of prediction for human actions. In one experiment, for example, the STS
response was enhanced for actions that were unpredicted for two different reasons:
reaching for empty space next to a target (which is an inefficient or failed action), or
reaching for a previously nonpreferred object (which is unpredicted relative to an
inferred goal; Carter et al., 2011; see also Bubic et al., 2009). It would be interesting to
test whether these two kinds of errors are associated with spatially distinct sources of
functional connectivity to the STS.
Using predictive coding to test the neural computations underlying
theory of mind
The framework of predictive coding offers a new opportunity to study the neural
representations of others’ actions and thoughts, using new experimental designs. The
necessary logic has been developed in repetition suppression experiments (GrillSpector et al., 2006). Complex stimuli elicit responses in many different brain regions
simultaneously, making it hard to dissociate the representational and computational
contributions of different brain regions. Consequently, in higher level vision, repetition
suppression has been used to differentiate the stimulus dimensions or features
represented in multiple co-activated regions. For example, although both the FFA and
the STS face area show repetition suppression when the identity of a face is repeated,
only a more anterior STS region shows a reduced response when the emotional
expression is repeated across different faces (Winston et al., 2004).
Looking for prediction error offers a generalized, and more flexible, version of repetition
suppression studies; critically, it only requires that a stimulus be surprising along some
dimension, without having to repeat the stimulus. This flexibility is particularly
advantageous for studies using naturalistic or social stimuli, which can be hard to repeat
without also invoking repetition of a number of confounding, but unrelated
representations — for example, words, syntactic structure, or faces. Exact repetitions of
complex stimuli can be unnatural or pragmatically odd, which may especially limit the
137 of 179
ability to study repetition suppression in young or special populations. By contrast, the
distribution of observed error signals could reveal both which neural populations or
regions are coding the relevant dimensions and features, and what the sources of
predictions are.
Finally, and perhaps most importantly, this framework may enrich theorizing about
neuroimaging results in social cognitive neuroscience. One of the key challenges facing
social cognitive neuroscience is that the richness of the data often surpasses the
precision of the theories. This proves to be a problem both for interpreting the data —
inverse inferences are very rarely well-constrained enough to be compelling, despite
their role in theory building — and for designing new hypotheses and experiments.
Increased response in a brain region has been argued to indicate both that the stimulus
carries many relevant features to a region and that the stimulus was harder to process
or a less good “fit” to the region; this problem is exacerbated when trying to interpret
different neural patterns across groups (i.e. special populations). If we can begin to
break down (a) what kinds of predictions a region makes, (b) what kind of information
directs those predictions, and (c) what constitutes an error, it may be possible to
formulate much more specific hypotheses about the computations, and information flow,
that underly human theory of mind.
In sum, we find a predictive coding approach to theory of mind promising. There is
extensive evidence of a key signature of predictive coding, in fMRI studies of theory of
mind: reduced responses to expected stimuli. Existing data also provide hints of other,
more distinctive signatures of predictive coding. Future experiments designed to more
directly test the predictions and errors represented in different brain regions may
provide an important new window on the neural computations underlying theory of mind.
Acknowledgements
The authors thank Amy Skerry, Hilary Richardson, Todd Thompson, and Nancy
Kanwisher for comments and discussion. The authors gratefully acknowledge support of
this project by an NSF Graduate Research Fellowship (#0645960 to JKH) and an NSF
CAREER award (#095518), NIH (1R01 MH096914-01A1), and the Packard Foundation
(to RS).
138 of 179
Chapter 6.
General Discussion
Summary
Thinking about another’s thoughts increases metabolic activity in a reliable set of
cortical regions, often called the theory of mind network, which includes bilateral
temporo-parietal junction (right and left TPJ), right superior temporal sulcus (right STS),
medial precuneus (PC), and medial prefrontal cortex (MPFC). These regions show
robust and systematic responses to a variety of stimuli that invoke a mental state
attribution (Fletcher et al., 1995; Goel et al., 1995; Gallagher et al., 2000, 2002; Mitchell
et al., 2002; Saxe & Kanwisher, 2003; Perner, Aichhorn, Kronbichler, Wolfgang, &
Laddurner, 2006; Gobbini, Koralek, Bryan, Montgomery & Haxby, 2007; Van Overwalle,
2009; Walter et al., 2010), and respond robustly to mental state information in
comparison both to other social information, including appearance, cultural background,
and subjective sensation such as exhaustion or hunger (Saxe & Powell, 2006; Saxe &
Wexler, 2005), and to non-mental representations, including misleading signs, out-ofdate maps, and temporary changes in reality (Saxe & Kanwisher, 2003; Perner, et al.,
2006). While many studies have investigated the selectivity and domain specificity of
the brain regions for theory of mind (e.g., Aichhorn et al., 2009; Saxe & Kanwisher,
2003), a key challenge is to go beyond basic involvement to determine the
computational roles of these regions: which aspects of people’s beliefs and intentions
are represented, or made explicit, in these brain regions, and how are they coded?
The goal of this thesis is to lay the theoretical and empirical groundwork for
systematically characterizing the representations and computations in mental state
inference. Building on prior work, which has for the most part focused on where in the
brain mental state reasoning occurs, the research here investigates how neural
populations encode concepts underlying mental state inference. In Chapter 1, I
reviewed existing literature on the neuroscience of mental state inference, to argue that
an important next step in understanding the neural basis of social cognition is to
understand the neural representations supported by “social” brain regions. In Chapters
2, 3, and 4, I presented seven empirical studies that demonstrate the successful use of
modern neuroimaging techniques to find abstract and behaviorally relevant neural
representations inside ToM brain regions, and begin to characterize the content of these
representations. In Chapter 5, I offered a novel interpretation of the existing literature
based on a predicative coding model of the brain. Together, this work takes a first step
in characterizing the content and format of abstract, behaviorally relevant features of
mental state inferences inside cortical regions that support social cognition.
139 of 179
Where We Are
Content in theory of mind brain regions
One of the major accomplishments of this thesis is to provide a first characterization of
the content represented by brain regions that support mental state inference. In three
chapters, and seven empirical studies, I present evidence that the neural code in brain
regions supporting social cognition is sensitive to difference types of mental state
representations, rather than simply distinguishing mental state atrributions from other
types of information.
In Chapter 2, I show that brain regions supporting social cognition distinguish between
mental state representations, that this code is detectable using fMRI, and that these
neural representations are behaviorally relevant. With different stimuli, paradigms, and
participants, I find converging results across three experiments: in neurotypical (NT)
adults, stories about intentional versus accidental harms elicited spatially distinct
patterns of response within the right temporo-parietal junction (RTPJ). Moreover, this
neural response mirrored behavioral judgments: individuals whose patterns in RTPJ
were more distinct also made a larger distinction between intentional and accidental
harms in their moral judgments. Notably, in the fourth experiment, this pattern was
absent in adults with autism spectrum disorders. Adults with ASD are characterized by
persistent social difficulty, including in using mental state information to reason about
the actions of others. Mirroring group-level differences in behavior, we found no neural
distinction between accidental and intentional harms in pattern or mean signal in adults
with ASD. These data suggest that one aspect of the role of RTPJ is to make explicit, in
the population response of its neurons, aspects of beliefs that are most relevant for
inference and decision-making — for example, encoding the intent behind a harmful act.
In Chapter 3, I find that information about the epistemic history of other’s beliefs is
represented in RTPJ, and that these neural representations are robust to profound
differences in developmental history. I find that the TPJ encodes information about the
source (specifically, the sensory modality) of others’ mental states, and that these
representations are preserved in the RTPJ of congenitally blind adults. The existence of
these representations in congenitally blind adults, who have no first-person experience
with sight, provides evidence that representations of other’s epistemic history emerge
even in the absence of relevant first-person perceptual experiences. Moreover, it
suggests that the coding of this information must be relatively abstract, rather than tied
to a single type of input or grounded in a specific sensory domain. Together, these
results suggest that the temporo-parietal junction contains explicit, abstract
representations of another feature mental states: the perceptual source.
In Chapter 4, I show that the neural representations of source modality found in
Chapter 3 are part of a more complex feature space of mental state representation, and
are coded along independent, continuous dimensions. I probed three epistemic features
of others’ beliefs (distinctions relevant to the source of another person’s belief), and one
salient but non-epistemic distinction (the valence of the person's resulting emotion). In
two experiments, I find that RTPJ and RMSTS contain complementary, continuous
neural codes which track the degree to which someone’s evidence supports their
140 of 179
beliefs, reliably distinguishing between good, reliable evidence and obscured,
ambiguous evidence; that RTPJ and LTPJ encode information about the modality of
other people’s evidence; and that the neural pattern in VMPFC codes continuous
information about the valence of other people’s experiences. These results suggest that
the cortical regions implicated in mental state inference contain neural representations
of epistemic and emotional features of others’s beliefs, and that these features are
represented along continuous, abstract dimensions.
The data in these chapters suggest that the brain regions implicated in social cognition
contain explicit neural representations of abstract aspects of others’ beliefs, including
epistemic and emotional features, and that these features are represented in a
continuous, abstract feature space. The neural code in RTPJ (a region implicated in all
seven studies) in particular seems to contain information about the beliefs of others,
tracking information about the relationship of someone’s beliefs to their actions (intent),
the reliability of the evidence they based their beliefs on (evidence quality), and the
source of their information (modality). Right STS was also found to contain
representations that support tracking the epistemic features of others’ beliefs,
specifically the quality of others’ evidence, whereas Left TPJ contains information about
the source of someone’s belief — in particular, the modality through which the evidence
was acquired. Finally, VMPFC (unlike RTPJ), contains continuous representations about
how people feel, based on the beliefs they hold.
Finally, in Chapter 5, I take a step back to consider how we might interpret the
activation observed in each of these regions. I propose that an existing framework of
neural coding, predictive coding, can provide a unifying framework for interpreting —
and generating new hypotheses about — the kinds of representations we find in social
brain regions. Predictive coding models posit that the brain generates continuous
predictions of upcoming events, and then adjusts these predictions by computing an
error signal which tracks deviations between predicted and observed events. Human
observers appear to formulate predictions for other people’s movements, actions,
beliefs, preferences and behaviors, based on relatively abstract internal models of
people’s bodies, minds, and personalities. I argue that these predictions are reflected in
multiple brain regions, including RSTS, RTPJ, and MPFC. I observe that across a
variety of tasks and studies, the RSTS responds more to observed actions that are
unpredicted; the RTPJ appears to exhibit a heightened response to unpredicted
information about others’ mental states; and MPFC responds more to unpredicted
information about a person’s preferences or traits. In sum, there is extensive evidence
of a key signature of predictive coding, in fMRI studies of theory of mind: reduced
responses to expected stimuli. I propose that future experiments designed to more
directly test the predictions and errors represented in different brain regions may
provide an important new window on the neural computations underlying theory of mind.
A feature-based approach to mental state representations
This work has largely looked for “features” of mental states — is a particular mental
state representation “intentional” or “accidental”, “reliable” or “unreliable”, “happy” or
“sad? I find that theory of mind brain regions, and in particular, RTPJ, represent features
141 of 179
of others’ beliefs — in particular, features of the relationship between the content of
someone’s beliefs and other relevant information, such as the agent’s actions (intent),
and the evidence available in the world (source modality and evidence quality).
Interestingly, the one feature that I find outside of RTPJ (valence), is not a feature of a
relationship between a belief and another piece of information, but rather an outcome of
an agent having had a belief. Based on these data, an intriguing hypothesis is that one
of the main functions of lateral ToM brain regions, and perhaps the RTPJ specifically, is
to track relationships between beliefs and other concepts, while medial brain regions
track information about agents and individuals. If this is the case, we might expect that
information qualifying the relationship between an agent and their belief, such as
confidence (guesses versus believes), factivity (knows versus thinks), negation
(believes versus doubts), or qualifying relationships between multiple agents and beliefs
(agrees versus disagrees) would also be found in RTPJ, while information about agents,
such as reliability and trustworthiness, would be found in MPFC. If this is correct, a key
role of reasoning about the minds of the others may the ability to track the relationships
between pieces of information, such agents, attitudes, actions, and real-world evidence,
and one role of theory of mind brain may be to track these relationships.
However, a limitation to this work is that, so far, I have looked only for specific types of
features: the relationships between belief content and other components of mental state
attributions, rather than features of, for example, specific agents or belief content. How
can we move from testing two or three features to genuinely looking for a compositional,
flexible, feature space?
First, we can look for features of other parts of mental state attributions, for example
features of the agents holding the beliefs, and features of the belief content. Previous
work suggests that features of the agent might be represented in MPFC (WhitfieldGabrieli et al., 2011; Moran et al., 2011a; Saxe, Moran, Scholz, & Gabrieli, 2006a;
Krienen, Tu, & Buckner, 2010) and anterior temporal lobe (Zanh et al., 2007; Anzellotti,
et al., 2013). Interestingly, however, representations about the content of others’ belief
may not be represented by regions classically associated with mental state inference.
For example, we can reason about many different properties of the proposition The Red
Sox will win the World Series, (e.g. our confidence and emotional response, or its
grammaticality, tense, and likelihood of being true) without having to consider the
information in the context of someone else’s mind. It is possible therefore, is that the
content of mental state representations is not represented in regions of the brain that
specifically support mental state inference, but rather in regions, such as language
regions, that more generally support reasoning about propositions (e.g. Friederici,
2002). If that is the case, we would expect that features of belief content would appear
in the same regions that support other semantic distinctions and processing of
propositional content (Fairhall et al, 2013; Mahon & Caramazza, 2010; Binney et al,
2010; Visser et al. 2012).
A second approach is to use “data-driven” analyses to find potential neural feature
spaces. This approach, which has been applied recently in a number of domains, such
as object recognition (Kriegeskorte, et al., 2008), audition (Norman-Haignere et al.,
2014) and action observation (Deen & Saxe, in prep), takes the neural response to a
142 of 179
diverse and relatively uncontrolled set of stimuli, and asks which classes of stimuli elicit
similar responses, using a variety of dimensionality reduction techniques. The
advantage of this approach is that it can find both a small number of dimensions that
explain a good portion of the variance in their stimuli set, and possibly find new, unhypothesized dimensions in the data. These dimensions can then be used to generate
new hypotheses, to be tested in follow-up experiments. This technique may reveal
important dimensions in the neural representations of social cognition. However, it also
has clear drawbacks. For example, when a dimension emerges that is not interpretable,
it is difficult to know if the researcher picked the wrong clustering methods, or if the
brain is grouping information along dimensions that human minds haven’t yet thought
of. Moreover, it can be difficult to measure whether the chosen stimulus set exhaustively
covers the cognitive or neural domain (and thus whether the features discovered truly
explain most of the variance in a given domain). Nevertheless, a combination of datadrive approaches and hypothesis-driven approaches, may be the key next steps in
characterizing the broader feature space of mental state representation.
Beyond Features
Feature-based approaches have proven invaluable in a number of cognitive and neural
domains, including vision (e.g. Shepard, 1980; Serre et al, 2005; Harrison et al, 2009),
memory, (e.g. Eichenbaum, 2004) and language (e.g. Naselaris et al, 2011), concept
representation (Mahon et al., 2010; Peelen et al., 2012). However, this approach also
has clear limitations. It is unlikely that mental state attributions are best described as
simply a list of features; rather, mental state attributions are likely representations with
internal structure (Baker at al, 2009, 2011; Davidson, 1963; Fodor 1981, 1987; Chen
2009), composed, for example, of an agent (the belief holder, e.g. John), the content of
the belief (e.g. it is raining outside), an attitude (the relationship between the agent and
the belief content, e.g. suspects), and perhaps in some cases, a representation of the
world outside of the agent’s beliefs (e.g., But it is actually sunny). These structures show
the hallmarks of being the product of a compositional, generative process (Chomsky,
1995): agents, attitudes, and beliefs can be combined in novel ways to form coherent
mental state representations (Snuffleupagus doubts that Red Sox will win the World
Series), and mental states can be embedded in one another (Snuffleupagus doubts that
Elmo believes that the Red Sox will win the World Series). Purely feature-based
approaches will fail to capture many of these dimensions, in particular hierarchical and
causal relationships.
Nevertheless, a feature-based approach has proven to be a fruitful first step in
characterizing the representations inside ToM brain regions. Recent computational work
has worked on combining basic “bags of features” approaches with structured
representations, which have increased predictiveness and explanatory power in a
number of domains (e.g. Kemp & Tenebaum 2008; Lazebnik et al, 2006). Given the
success of a feature-based approach here, and in much of higher-level cognitive
neuroscience, exploring models that use features to infer underlying structure may
prove invaluable to answering some of the harder questions about the cognitive and
neural representations of abstract concepts.
143 of 179
Continuous, abstract representations
One way to leverage information about features to infer more about the underlying
structure is to look past simple binary tags. Here, I ask whether the dimensions we
detect using feature-based approaches are binary or gradient. In Chapter 4, I show that
the probability of an item being categorized as one type of stimulus, based on neural
data, is predicted by independent behavioral ratings, suggesting that the brain may
represent these stimuli along continuous, abstract feature dimensions (Kriegeskorte,
2008; Kriegeskorte & Kievit, 2013), rather than with binary tags. Interesting open
questions are whether these regions contain both continuous and discrete dimensions
(for example, perhaps modality is coded discretely), and to what extent they are fully
combinatorial: does every possibly combination of features get represented?
Though these results just scratch the surface of what can be learned about the format of
representations in ToM regions, taken together, they are particularly striking because
they find evidence of neural representations of highly abstract conceptual features, and
thus add to the growing body of literature that has found neural representations of a
variety of high-level cognitive domains (e.g., Anzellotti et al., 2013; Hassabis et al.,
2013; Behrens, Hunt, & Rushworth, 2009; Izuma, Saito, & Sadato, 2008; Poore et al.,
2012; Correia et al., 2014).
Inputs to representations
A second way to leverage feature-based approaches to understand underlying structure
is to examine the information that serves as input to these features, and ask to what
extent the features are invariant to changes in that input. Finding that a distinction
between two features is represented in a neural code fails to tell us about its underlying
structure. For example, evidence that there is a neural distinction between knowing that
someone learned a fact by seeing it as opposed to by hearing it could indicate a
distinction between types of sensory simulations — a “re-experience” of that person’s
hearing or seeing. It could also map onto a more abstract, conceptual distinction, e.g., a
model of perceptual access that encodes that some information can be acquired
through one type of source, and not another. We can begin to distinguish between these
two different types of underlying structure by asking about the kind of information that
feeds these representations: is the input necessarily first-person sensory motor cues,
or, does it come from a broader range of sources? Here, I ask whether the features
found in ToM brain regions are tied to specific sources of information (for example,
sensory-motor input) by asking whether the existence of these representations depends
on having specific machinery or experiences.
In Chapter 3, I find that the RTPJ in blind adults contains distinct neural representations
of seeing and hearing. The developmental history that supports (or limits) a neural
representation can tell us information about its source. By asking whether a neural
representation has been changed in individuals who have different basic building
blocks, such as those with different sensory experiences, we can place important limits
on the possible input of the representation. By asking how blind individuals represent
others’ experiences of sight, for example, we can learn the extent to which theory of
mind depends on the observer having had similar sensations, experiences, beliefs, and
feelings as the target, and constrain our theories of mental state inference accordingly.
144 of 179
Finding robust representations of others’ experience of sight in the RTPJ of blind adults
suggests that, much like representations of value in pre-frontal cortex (e.g. O'Doherty et
al., 2001; Wallis et al., 2001), representation of emotion in MPFC (Skerry & Saxe, in
prep; Peelen et al., 2013), or representations of semantic concepts in language regions
(e.g. Simanova et al., 2014), the neural representation of modality found in RTPJ is not
tied to a single type of input, or grounded in a specific sensory domain. Instead, we
have learned that these regions must be able to take a wide variety of information as
their input (sensory cues, linguistics information, inference), and that the information
inside these regions must be represented in a feature space general enough to support
information across these sources.
An important limitation to this finding, however, is that although we find a difference
between seeing and hearing stories in the same neural area in both blind and sighted
adults, this does not guarantee that we are finding the same difference. In fact, though
the stimuli in Chapter 3 were written to manipulate only modality, they did not explicitly
code for quality of evidence — a dimension which, we discovered in Chapter 4, is also
represented in RTPJ. Therefore it is possible that while the RTPJ of sighted adults
independently represents both modality and quality, the RTPJ of blind adults represents
only quality, though this interpretation seems less likely in light of the specific stimuli
used in Chapter 3. The stimuli were matched across conditions to be minimal pairs
(e.g., he saw her face versus he heard her voice). Importantly, the stimuli often sidestepped the issue of evidence quality by explicitly stating the final outcome of the event,
rather than the evidence (e.g., John heard/saw the glass shatter, rather than John saw
glass shards or John heard a breaking sound), allowing little room to make inferences
about how well the evidence supported the conclusion. Moreover, the inference that
blind and sighted people have similar representations of sight is strengthened both by
blind people’s extensive knowledge of vision (Landau & Gleitman, 1985; Lenci, Baroni,
Cazzolli, & Marotta, 2013, Koster-Hale et al., in prep), and by our use of a priori (and
rather small) regions of interest. Finding a distinction between seeing and hearing
stories in both groups anywhere in the brain would be weak evidence that both groups
make the same distinction. However, finding this distinction in a constrained ROI,
chosen for its known role in mental state processing, makes it less likely that, by
coincidence, this region is involved in distinguishing e.g. between sensory simulations of
seeing and hearing events, in sighted people, and between familiar and unfamiliar
events, in blind people. Thus, though it is more parsimonious to infer that both groups
distinguish between these stimuli in the same region for the same reason, import future
work will need to characterize to what extent these representations are truly shared, and
where the limitations in blind people’s representations in sight are.
Replication and generalization
One the challenges of trying to characterize abstract and complex cognitive processes
using fMRI, is that the data are often too limited to draw meaningful inferences about
true conceptual structure. Even ignoring fundamental limitations in the technique (the
resolution of the data acquired through human neuroimaging is far too low to ask
questions about how single neurons (or even populations of cells) code information —
each data point blurs two seconds of activity in hundreds of thousands of neurons)
145 of 179
finding that a distinction exists in a neural population often tells us little about what part
of the distinction is being represented. Such ambiguity is due in part to a vast array of
potential dimensions that any empirically observable distinction could map on to. There
may be hidden dimensions along which complex stimuli grouped into separate
conditions differ and that it is these differences, rather than the intended manipulation,
that lead to differential brain activity.
One way to tackle that problem is by asking whether the distinction generalizes across
different manipulations of the distinctions. If a distinction is observed across multiple
ways of drawing the conceptual distinction, the observed difference in stimuli is more
likely to map onto the intended manipulation, rather than superficial properties of the
stimulus or task. Here, I find that the three key findings observed in this thesis replicate
across a diverse range of stimuli and tasks. In Chapter 2, I replicate the first key result
in three different experiments, showing that RTPJ distinguishes between accidental and
intentional harm across stimuli, task, and participants. Chapter 3, I replicate the second
key result across participant populations, showing that bilateral TPJ codes information
about source, and in Chapter 4 offer a third replication, using a new design, stimuli, and
participant pool. Chapter 4 also contains a replication the third key result, showing that
RTPJ codes information about the quality of others’ evidence, across task, stimuli,
participants, and critically, across both specific perceptual of good versus bad evidence,
and more abstract ones. (Interestingly, the representations in RMSTS do not generalize,
suggesting that the distinction in RMSTS is not precisely good versus bad evidence, but
rather a more specific dimension that was manipulated, such as quality of sensory
information.) Taken together, these data demonstrate that it is possible to find abstract,
behaviorally relevant features of mental state inferences inside cortical regions that
support social cognition, and have taken a first step in characterizing their content and
structure.
What next?
Information between regions, and between networks
Though this thesis largely focus on the content inside ToM brain regions, the results hint
at differences in the information available in each ToM region. The data here suggest
that lateral regions, including RTPJ, RMSTS, and LTPJ, represents features of beliefs;
and in particular the epistemic features of beliefs, including the quality and modality of
the evidence that gives rise to other people’s beliefs, while MPFC may track features of
individual agents, such as emotional response. These results converge with recent work
suggesting a division of labor between regions. Lateral and medial regions show
differences in overall activation profiles (e.g. Mitchell et al., 2006; Jacques. Conway,
Lowder, & Cabeza, 2011; Krienen et al., 2010), and contain different information in their
neural code. MVPA has shown that RTPJ contains information about future social
actions (Carter, Bowling, Reeck, & Huettel, 2012), social groups versus individuals
(Contreras, Schirmer, Banaji, & Mitchell, 2013), and intent when causing harm (Chapter
2, Koster-Hale et al., 2013), and the reliability of others’ cooperation (Behrens, Hunt,
Woolrich, & Rushworth, 2008). In contrast, MPFC has been shown to track information
about long-term personality traits (Hassabis et al., 2013) and emotional states (Kassam,
146 of 179
Markey, Cherkassky, Loewenstein, & Just, 2013; Peelen et al., 2010). Knowing how
regions are different is just one step in understanding how the brain works to give rise to
many complexities of social cognition. Future work needs to more systematically
characterize not only the division of labor between these regions, but also their
interactions with each other, and with regions in other networks.
Development
This thesis adds to a growing body of work showing that parts of the human brain are
selective for social cognition. A key question, then, is what triggers the specialization of
these brain regions: hard-wired maturation or experience during development? One
possibility is that some number of cognitive functions are implemented by innately
specified brain regions (Fodor 1983). These (potentially) modular brain regions would
allow rapid and efficient computation of evolutionarily significant cognitive processes, for
example, vision. On this view, the association between a selective functional profile,
such as increased response to mental state inference, and a specific brain region, such
as the RTPJ, would provide evidence that mental state inference is subserved by an
innate cognitive module. An alternative possibility is that specific brain regions acquire
highly specialized functions through experience. In particular, extensive early
experience may drive the organization of cortex, so that populations of neurons adopt
specific functions. For example, behavioral training in monkeys increases neuronal
selectivity in cortex recruited for object recognition (Baker et al. 2002; Sigala &
Logothetis 2002; Freedman et al. 2006; Cox & DiCarlo 2008). If these populations of
neurons are spatially contiguous, functionally coherent patches of cortex may be visible
using techniques like fMRI. One such example is the visual word form area, a region in
human temporal cortex that responds selectively to words in familiar (but not unfamiliar)
alphabets (Baker et al. 2007). If specialized brain regions can be organized by intense
early experience, then RTPJ may acquire its highly specific response to other minds
over the course of childhood experiences. Another important line of research will be to
test these hypotheses.
One way to test this hypothesis is to study the neural mechanism of ToM in individuals
who have had atypical developmental histories with regard to theory of mind. For
example, prior evidence suggests that early development of ToM depends on exposure
to language and to conversation about the mind (e.g. Milligan et al. 2007; Moeller &
Schick 2006; de Villiers 2005). D/deaf 15 children born to non-signing parents often have
delayed access to language, including language about mental states compared, to D/
deaf children born to signing parents. As a consequence, they exhibit delayed
performance on standard measures of ToM reasoning (Figueras-Costa & Harris 2001;
Meristo et al. 2007; Peterson & Wellman 2010; Peterson & Siegal 1999; Schick et al.
2007; Woolfe & Want 2002). However, later in childhood, these children appear to catch
up in ToM performance (Peterson & Wellman 2010; Schick et al. 2007). Since most d/
Deaf children are otherwise typically developing, but vary in their early language
15 Deaf
with an uppercase "D" usually refers to adults and children who share the use of American Sign
Language and Deaf culture, while a lowercase "d" is usually an audiological description of a person's
hearing level.
147 of 179
acquisition, they provide a unique window into the effects of early language experience
on theory of mind. In ongoing work, I am asking whether this difference in language
experience, and subsequent delay in ToM development, leave traces on the neural
basis of ToM.
Links to behavior: Content and structure of mental state inference
Finally, a core goal of understanding the neural basis of social cognition is to
understand how it gives rise to cognition. Perhaps not surprisingly, the questions that
have been asked about the neural basis of theory of mind reflect many of the questions
that have been asked behaviorally. Researchers have examined the innateness and
domain specificity of theory of mind (e.g. Carey & Spelke, 1996; Apperley et al, 2005),
and explored how mental state inference interacts with, and relies on, other cognitive
domains, such as language (e.g. Apperly et al., 2004, 2005; Astington & Baird (2005);
Milligan et al. 2007; Moeller & Schick 2006; de Villiers 2005) and executive function
(Ozonoff et al., 1991; Perner & Lang, 1999; Schneider et al., 2014). Others have tried to
characterize the prerequists and development of theory of mind (e.g. Wellman et al.,
2001; Saxe, 2013, Baillargeon et al., 2010), the cues we use when we infer that other
minds exist (e.g. Stich and Nichols, 2003), and the types of domain-general input that
feed mental state inferences (e.g. Apperly, 2014). Thus, while much work has been
done to ask what is (and isn't) in the domain of theory of mind, how it develops, and
where it sits in relation to the rest of the cognitive system, few behavioral approaches
ask about the content inside theory of mind: what the representational or computational
aspects of the mental representation itself are; what the input data, transformation of
that data, and subsequent outputs might be. Here, cognitive neuroscience may be able
to inform theories of cognition. Understanding how the brain represents complex and
abstract information to support mental state inference, and social cognition more
generally, can lay the groundwork for studying the cognitive structure of mental state
inference.
Beyond neuroimaging
However, neuroimaging is cumbersome, expensive, and limited. Many basic questions
about ToM cannot be addressed with neuroimaging. For example, a scientific theory of
how humans understand other minds should address questions like: “When and why do
we (spontaneously) seek to understand another’s thoughts?”, “How do we figure out the
actual content of someone else’s thoughts (i.e. what they are thinking) from specific
cues?", “How do we choose whether or not to incorporate others’ thoughts into our own
decisions and actions?”, and “Why do we care emotionally about others’ thoughts and
feelings?” None of these questions have yet been approached using neuroimaging, and
may pose much bigger challenges than the questions I have addressed so far.
Contemporary neuroimaging technology does not even allow us to address many
fundamental questions about the neural mechanisms of ToM. Existing tools are
extremely slow and blurry, by comparison to the speed and precision of neural
computation: they cannot decipher what is the input of a region, how that input is
transformed, or where the output from that region is sent, during mental state inference.
148 of 179
If our final horizon is a complete theory of how the brain allows us to understand other
minds, we will need to make dramatic progress on (1) the “psychophysics” of ToM in
adulthood, to allow precise quantitative measurements of people’s use of ToM; (2) a
computational model of ToM that is sufficiently explicit to make quantitative predictions
about adult judgments (e.g. Baker, Saxe, & Tenenbaum, 2011); and (3) a mechanism of
how neurons and networks of neurons might implement that computational model, by
sequentially transforming patterns of input into patterns of output that make different
information explicit. That horizon is still far away. However, the fact that we can give a
characterization of some of the boundary conditions that a successful account of ToM
needs to meet is part of what makes this such an exciting time to participate in the
cognitive neuroscience of understanding other minds.
149 of 179
150 of 179
References
Adams, R., Jr., Rule, N.O., Franklin Jr, R.G., Wang, E., Stevenson, M.T., Yoshikawa, S. (2010). Crosscultural reading the mind in the eyes: An fMRI investigation. Journal of Cognitive Neuroscience 22(1),
97-108.
Adolphs, R.R. (2001). The neurobiology of social cognition. Current Opinion in Neurobiology.
Adolphs, R.R. (2009). The social brain: neural basis of social knowledge. Annual review of psychology 60,
693.
Adolphs, R.R. (2010). Conceptual Challenges and Directions for Social Neuroscience. Neuron, 65(6),
752-767. doi:10.1016/j.neuron.2010.03.006
Adolphs, R.R. (2010). Conceptual challenges and directions for social neuroscience. Neuron, 65(6):
752-67.
Aichhorn, M. et al. (2006). Do visual perspective tasks need theory of mind? NeuroImage, 30(3): 1059-68.
Aichhorn, M. M., Perner, J., Weiss, B., Kronbichler, M., Staffen, W., & Ladurner, G. (2009). Temporoparietal junction activity in theory-of-mind tasks: falseness, beliefs, or attention. Journal of Cognitive
Neuroscience, 21(6), 1179-1192. doi:10.1162/jocn.2009.21082
Aichhorn, M., Perner, J., Kronbichler, M., Staffen, W., & Ladurner, G. (2009). Temporo-parietal junction
activity in theory-of-mind tasks: Falseness, beliefs, or attention. Journal of Cognitive Neuroscience
21(6): 1179-92
Aichhorn, M., Perner, J., Weiss, B., Kronbichler, M., Staffen, W., and Ladurner, G. (2009). Temporoparietal junction activity in theory-of-mind tasks: falseness, beliefs, or attention. J Cogn Neurosci,
21(6), 1179-1192.
Aikhenvald, A. Y. (2004). Evidentiality. Oxford University Press.
Alink, A., Schwiedrzik, C. M., Kohler, A., Singer, W., and Muckli, L. (2010). Stimulus predictability reduces
responses in primary visual cortex. J Neurosci, 30(8), 2960-2966.
Allison, T., Puce, A., and McCarthy, G. (2002). Category-sensitive excitatory and inhibitory processes in
human extrastriate cortex. J Neurophysiol, 88(5), 2864-2868.
American Psychiatric Association, & American Psychiatric Association Task Force on DSM-IV. (2000).
Diagnostic and statistical manual of mental disorders : DSM-IV-TR. Washington, DC: American
Psychiatric Association.
Ames, D.L., Jenkins, A.C., Banaji, M.R., & Mitchell, J.P. (2008). Taking another person's perspective
increases self-referential neural processing. Psychological Science 19(7): 642-4.
Anzellotti, S., Fairhall, S. L., & Caramazza, A. (2013). Decoding Representations of Face Identity That are
Tolerant to Rotation. Cerebral Cortex. doi:10.1093/cercor/bht046
Apperly, I. & Butterfill, S. (2009). Do humans have two systems to track beliefs and belief-like states?.
Psychological Review 116(4): 953-70.
Apperly, I. A. (2008). Beyond Simulation-Theory and Theory-Theory: why social cognitive neuroscience
should use its own concepts to study "theory of mind". Cognition, 107(1), 266-283. doi:10.1016/
j.cognition.2007.07.019
Apperly, I. A., Samson, D., Chiavarino, C., Bickerton, W. L., & Humphreys, G. W. (2007). Testing the
domain-specificity of a theory of mind deficit in brain-injured patients: Evidence for consistent
performance on non-verbal,'reality-unknown' false belief and false photograph tasks. Cognition,
103(2), 300-321.
151 of 179
Apperly, I.A., Samson, D., & Humphreys, G.W. (2005). Domain-specificity and theory of mind: evaluating
neuropsychological evidence. Trends in Cognitive Sciences 9(12): 572-7.
Arzy, S., Thut, G., Mohr, C., Michel, C.M., & Blanke, O. (2006). Neural basis of embodiment: distinct
contributions of temporoparietal junction and extrastriate body area. Journal of Neuroscience 26(31):
8074-81.
Averbeck, B. B., Latham, P. E., & Pouget, A. (2006). Neural correlations, population coding and
computation. Nature Reviews Neuroscience, 7(5), 358-366. doi:10.1038/nrn1888
Aydin, â., & Ceci, S. J. (2009). Evidentiality and suggestibility: A new research venue - Aydin - 2009 - New
Directions for Child and Adolescent Development - Wiley Online Library. New Directions for Child and
Adolescent ƒ.
Baez, S., Rattazzi, A., Gonzalez-Gadea, M. L., Torralva, T., Vigliecca, N. S., Decety, J., et al. (2012).
Integrating intention and context: assessing social cognition in adults with Asperger syndrome.
Frontiers in Human Neuroscience, 6, 302. doi:10.3389/fnhum.2012.00302
Baird, J., & Astington, J. W. (2004). The role of mental state understanding in the development of moral
cognition and moral action. New Directions for Child and Adolescent Development, 2004(103), 37-49.
Baillargeon, R., Scott, R. M., & He, Z. (2010). False-belief understanding in infants. Trends in cognitive
sciences, 14(3), 110-118.
Baker, C. L., Saxe, R.R., and Tenenbaum, J. B. (2009). Action understanding as inverse planning.
Cognition, 113(3), 329-349.
Baker, C. L., Saxe, R.R., & Tenenbaum, J. B. (2011). Bayesian theory of mind: Modeling joint belief-desire
attribution. In Proceedings of the thirty-second annual conference of the cognitive science society (pp.
2469-2474).
Bargh, J.A., & Thein, R.D. (1985). Individual construct accessibility, person memory, and the recalljudgment link: The case of information overload. Journal of Personality and Social Psychology 49(5):
1129.
Barlow, H. B. (1961). Possible principles underlying the transformation of sensory messages. Sensory
communication, 217-234.
Baron-Cohen, S. (1995). Mindblindness: As essay on autism and theory of mind. Recherche.
Baron-Cohen, S. (2001). Theory of mind in normal development and autism. Prisme.
Baron-Cohen, S. L., MO'Riordan. (1999). A new test of social sensitivity: Detection of faux pas in normal
children and children with Asperger syndrome. Journal of Autism and Developmental Disorders.
Baron-Cohen, S., & Wheelwright, S. (2004). The empathy quotient: an investigation of adults with
Asperger syndrome or high functioning autism, and normal sex differences. Journal of Autism and
Developmental Disorders 34(2): 163-75.
Baron-Cohen, S., O'Riordan, M., Jones, R., Stone, V., & Plaisted, K. (1999). A new test of social
sensitivity: Detection of faux pas in normal children and children with Asperger syndrome. Journal of
Autism and Developmental Disorders 29, 407-18.
Baron-Cohen, S., Wheelwright, S., Hill, J., Raste, Y., & Plumb, I. (2001). The 'Reading the mind in the
eyes' test revised version: A study with normal adults, and adults with asperger syndrome or highfunctioning autism. Journal of Child Psychology and Psychiatry 42(2): 241-51.
Battelli, L., Cavanagh, P., & Thornton, I. M. (2003). Perception of biological motion in parietal patients.
Neuropsychologia, 41(13), 1808-1816.
Bayer, H.M., and Glimcher, P.W. (2005). Midbrain Dopamine Neurons Encode a Quantitative Reward
Prediction Error Signal. Neuron 47, 13-13.
152 of 179
Bedny, M., & Saxe, R.R. (2012). Insights into the origins of knowledge from the cognitive neuroscience of
blindness. Cognitive Neuropsychology, 29(1-2), 56-84. doi:10.1080/02643294.2012.713342
Bedny, M., McGill, M., & Thompson-Schill, S. L. (2008). Semantic Adaptation and Competition during
Word Comprehension. Cerebral Cortex, 18(11), 2574-2585. doi:10.1093/cercor/bhn018
Bedny, M., Pascual-Leone, A., & Saxe, R.R. (2009). Growing up blind does not change the neural bases
of Theory of Mind. Proceedings of the National Academy of Sciences of the United States of America,
106(27), 11312-11317. doi:10.1073/pnas.0900010106
Behrens, T. E. J., Hunt, L. T., & Rushworth, M. F. S. (2009). The Computation of Social Behavior. Science,
324(5931), 1160-1164. doi:10.1126/science.1169694
Behrens, T. E. J., Hunt, L. T., Woolrich, M. W., & Rushworth, M. F. S. (2008). Associative learning of social
value. Nature, 456(7219), 245-249. doi:10.1038/nature07538
Berthoz, S., Armony, J. L., Blair, R. J. R., and Dolan, R. J. (2002). An fMRI study of intentional and
unintentional (embarrassing) violations of social norms. Brain, 125(8), 1696-1708.
Bhatt, M., Lohrenz, T., Camerer, C.F., & Montague, P.R. (2010). Neural signatures of strategic types in a
two-person bargaining game. In Proceedings of the National Academy of Sciences 107(46): 19720-5
Bialek, W., Rieke, F., Van Steveninck, R.R. R., & Warland, D. (1991). Reading a neural code. Science,
252(5014), 1854-1857.
Bigelow, A. (1991). Hiding in blind and sighted children. Development and Psychopathology, (3), 301-310.
Bigelow, A. (1992). The development of blind children's ability to predict what another sees, (86),
181-184.
Binney, R. J., Embleton, K. V., Jefferies, E., Parker, G. J., & Ralph, M. A. L. (2010). The ventral and
inferolateral aspects of the anterior temporal lobe are crucial in semantic memory: evidence from a
novel direct comparison of distortion-corrected fMRI, rTMS, and semantic dementia. Cerebral Cortex,
20(11), 2728-2738.
Blair, J. (1996, October 30). Brief Report: Morality in the Autistic Child. Journal of Autism and
Developmental Disorders. Retrieved May 30, 2011, from
Blake, R., and Shiffrar, M. (2007). Perception of human motion. Annu Rev Psychol 58, 47-73.
Blakemore, S. J. (2005). Somatosensory activations during the observation of touch and a case of visiontouch synaesthesia. Brain, 128(7), 1571-1583. doi:10.1093/brain/awh500
Blanke, O., & Arzy, S. (2005). The out-of-body experience: disturbed self-processing at the temporoparietal junction. Neuroscientist 11(1): 16-24.
Blanke, O., Mohr, C., Michel, C.M., Pascual-Leone, A., Brugger, P., Seeck, M., et al. (2005). Linking outof-body experience and self processing to mental own-body imagery at the temporoparietal junction.
Journal of Neuroscience 25(3): 550-7.
Bloom, P., & German, T. (2000). Two reasons to abandon the false belief task as a test of theory of mind.
Cognition, 77(1), B25-B31.
Bonda, E., Petrides, M., Ostry, D., & Evans, A. (1996). Specific involvement of human parietal systems
and the amygdala in the perception of biological motion. J Neurosci, 16(11), 3737-3744.
Bortoletto, M., Mattingley, J. B., & Cunnington, R. (2011). Action intentions modulate visual processing
during action perception. Neuropsychologia, 49(7), 2097-2104.
Botvinick, M., Jha, A. P., Bylsma, L. M., Fabian, S. A., Solomon, P. E., & Prkachin, K. M. (2005). Viewing
facial expressions of pain engages cortical areas involved in the direct experience of pain.
NeuroImage, 25(1), 312-319. doi:10.1016/j.neuroimage.2004.11.043
153 of 179
Bradley, M.M., Sabatinelli, D., Lang, P.J., Fitzsimmons, J.R., King, W., and Desai, P. (2003). Activation of
the visual cortex in motivated attention. Behav Neurosci 117, 369-380.
Brambring, M., & Asbrock, D. (2010). Validity of false belief tasks in blind children. Journal of Autism and
Developmental Disorders, 40(12), 1471-1484. doi:10.1007/s10803-010-1002-2
Brass, M., Schmitt, R.M., Spengler, S., and Gergely, G. (2007). Investigating Action Understanding:
Inferential Processes versus Action Simulation. Curr Biol 17, 2117-2121
Bray, S., Chang, C., & Hoeft, F. (2009). Applications of multivariate pattern classification analyses in
developmental neuroimaging of healthy and clinical populations. Frontiers in Human Neuroscience, 3.
doi:10.3389/neuro.09.032.2009
Brieber, S., Herpertz-Dahlmann, B., Fink, G. R., Kamp-Becker, I., Remschmidt, H., & Konrad, K. (2010).
Coherent motion processing in autism spectrum disorder (ASD): An fMRI study. Neuropsychologia,
48(6), 1644-1651. doi:10.1016/j.neuropsychologia.2010.02.007
Bright-Paul, A., Jarrold, C., & Wright, D. B. (2008). Theory-of-mind development influences suggestibility
and source monitoring. Developmental Psychology, 44(4), 1055-1068.
doi:10.1037/0012-1649.44.4.1055
Brown, E. C., & Brüne, M. (2012). The role of prediction in social neuroscience. Front Hum Neurosci, 6.
Bruneau E., Dufour N., & Saxe R. (2012) Social cognition in members of conflict groups: behavioural and
neural responses in Arabs, Israelis and South Americans to each other's misfortunes. Philosophical
Transactions of the Royal Society, London, B Biological Sciences 367(1589) 717-30.
Bruneau, E. G., & Saxe, R.R. (2012). Distinct roles of the 'shared pain' and 'theory of mind' networks in
processing others' emotional suffering. Neuropsychologia, 50(2), 219-231. doi:10.1016/
j.neuropsychologia.2011.11.008
Bruneau, E.G., & Saxe, R.R. (2010). Attitudes toward the outgroup are predicted by activity in the
precuneus in Arabs and Israelis. NeuroImage 52(4): 1704-11.
Bruneau, E.G., Pluta, A., & Saxe, R.R. (2011). Distinct roles of the 'shared pain' and 'theory of mind'
networks in processing others' emotional suffering. Neuropsychologia pp.1-13.
Bubic, A., Cramon, von, D.Y., Jacobsen, T., Schröger, E., and Schubotz, R.I. (2009). Violation of
expectation: neural correlates reflect bases of prediction. J Cogn Neurosci 21, 155-168.
Buckholtz, J. W., Asplund, C. L., Dux, P. E., Zald, D. H., Gore, J. C., Jones, O. D., & Marois, R. (2008).
The Neural Correlates of Third-Party Punishment. Neuron, 60(5), 930-940. doi:10.1016/j.neuron.
2008.10.016
Buhrmester, M., Kwang, T., & Gosling, S. D. (2011). Amazon's Mechanical Turk A New Source of
Inexpensive, Yet High-Quality, Data? Perspectives on Psychological Science, 6(1), 3-5. doi:
10.1177/1745691610393980
Bull Kovera, M., Park, R. C., & Penrod, S. D. (1992). Jurors' Perceptions of Eyewitness and Hearsay
Evidence. Minn L Rev, 76, 703.
Butts, D. A., Weng, C., Jin, J., Yeh, C.-I., Lesica, N. A., Alonso, J.-M., & Stanley, G. B. (2007). Temporal
precision in the neural code and the timescales of natural vision. Nature, 449(7158), 92-95. doi:
10.1038/nature06105
Bzdok, D., Schilbach, L., Vogeley, K., Schneider, K., Laird, A. R., Langner, R., & Eickhoff, S. B. (2012).
Parsing the neural correlates of moral cognition: ALE meta-analysis on morality, theory of mind, and
empathy. Brain Structure and Function. doi:10.1007/s00429-012-0380-y
Calder, A. J., Beaver, J. D., Winston, J. S., Dolan, R. J., Jenkins, R., Eger, E., & Henson, R. N. (2007).
Separate coding of different gaze directions in the superior temporal sulcus and inferior parietal
lobule. Curr Biol, 17(1), 20-25.
154 of 179
Calvo-Merino, B., Glaser, D. E., Grèzes, J., Passingham, R. E., & Haggard, P. (2005). Action observation
and acquired motor skills: an FMRI study with expert dancers. Cerebral Cortex, 15(8), 1243-1249.
doi:10.1093/cercor/bhi007
Calvo-Merino, B., Grèzes, J., Glaser, D. E., Passingham, R. E., & Haggard, P. (2006). Seeing or Doing?
Influence of Visual and Motor Familiarity in Action Observation. Current Biology, 16(19), 1905-1910.
doi:10.1016/j.cub.2006.07.065
Carey, S., & Spelke, E. (1996). Science and core knowledge. Philosophy of Science, 515-533.
Carrington, S. (2009). Are there theory of mind regions in the brain? A review of the neuroimaging
literature - Carrington - 2008 - Human Brain Mapping - Wiley Online Library. Human Brain Mapping.
Carrington, S.J., & Bailey, A.J. (2009). Are there theory of mind regions in the brain? A review of the
neuroimaging literature. Human Brain Mapping 30(8): 2313-35.
Carter, E.J., Hodgins, J.K., and Rakison, D.H. (2011). Exploring the neural correlates of goal-directed
action and intention understanding. NeuroImage 54, 1634-1642.
Carter, R. M., Bowling, D. L., Reeck, C., & Huettel, S. A. (2012). A Distinct Role of the Temporal-Parietal
Junction in Predicting Socially Guided Decisions. Science, 337(6090), 109-111. doi:10.1126/science.
1219681
Castelli, F., Happé, F. G. E., Frith, U., & Frith, C. (2000). Movement and mind: a functional imaging study
of perception and interpretation of complex intentional movement patterns. NeuroImage, 12(3),
314-325.
Castelli, F., Happé, F., Frith, U., & Frith, C. (2000). Movement and mind: a functional imaging study of
perception and interpretation of complex intentional movement patterns. NeuroImage 12(3): 314-25.
Chang, C.-C., & Lin, C.-J. (2011). LIBSVM: A library for support vector machines. ACM Transactions on
Intelligent Systems and Technology (TIST), 2(3), 27-27. doi:10.1145/1961189.1961199
Chang, L.J., and Sanfey, A.G. (2013). Great expectations: neural computations underlying the use of
social norms in decision-making. Soc Cogn Affect Neurosci 8, 277-284.
Chen, W. (2009). Understanding mental states in natural language. In Proceedings of the Eighth
International Conference on Computational Semantics (pp. 61-72). Association for Computational
Linguistics.
Chikazoe, J., Lee, D. H., Kriegeskorte, N., & Anderson, A. K. (2014). Population coding of affect across
stimuli, modalities and individuals. Nature Publishing Group, 1-11. doi:10.1038/nn.3749
Chomsky, N. (1995). The minimalist program (Vol. 28). Cambridge, MA: MIT press.
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science.
Behav Brain Sci 36, 181-204.
Clément, F., Koenig, M., & Harris, P. (2004). The Ontogenesis of Trust. Mind & Language, 19(4), 360-379.
doi:10.1111/j.0268-1064.2004.00263.x
Cloutier, J., Gabrieli, J.D.E., O'Young, D., & Ambady, N. (2011). An fMRI study of violations of social
expectations: When people are not who we expect them to be. NeuroImage 57(2): 583-8.
Cloutier, J., Gabrieli, J.D.E., O'Young, D., and Ambady, N. (2011). An fMRI study of violations of social
expectations: When people are not who we expect them to be. NeuroImage 57, 583-588.
Cohen, J.D., Perlstein, W.M., Braver, T.S., Nystrom, L.E., Noll, D.C., Jonides, J., and Smith, E.E. (1997).
Temporal dynamics of brain activation during a working memory task. Nature 386, 604-608.
Cohen, S.B., Wheelwright, S., & Hill, J. (2001). The 'reading the mind in the eyes' test revised version: A
study with normal adults, and adults with Asperger syndrome or high-functioning autism. Journal of
Child Psychology and Psychiatry. 42(2), 241-251.
155 of 179
Contreras, J. M., Schirmer, J., Banaji, M., & Mitchell, J. P. (2013). Common Brain Regions with Distinct
Patterns of Neural Responses during Mentalizing about Groups and Individuals. Journal of Cognitive
Neuroscience, 25(9), 1406-1417. doi:10.1162/jocn-a-00403
Corbetta, M., & Shulman, G.L. (2002). Control of goal-directed and stimulus-driven attention in the brain.
Nature Review Neuroscience 3: 201-15.
Corbetta, M., Kincade, J.M., Ollinger, J.M., McAvoy, M.P., & Shulman, G.L. (2000). Voluntary orienting is
dissociated from target detection in human posterior parietal cortex. Nature Neuroscience 3: 292-7.
Corbetta, M., Patel, G., & Shulman, G.L. (2008). The reorienting system of the human brain: from
environment to theory of mind. Neuron 58(3): 306-24.
Coricelli, G., & Nagel, R. (2009). Neural correlates of depth of strategic reasoning in medial prefrontal
cortex. Proceedings of the National Academy of Sciences 106(23): 9163-8.
Correia, J., Formisano, E., Valente, G., Hausfeld, L., Jansma, B., & Bonte, M. (2014). Brain-based
translation: fMRI decoding of spoken words in bilinguals reveals language-independent semantic
representations in anterior temporal lobe. The Journal of Neuroscience : the Official Journal of the
Society for Neuroscience, 34(1), 332-338. doi:10.1523/JNEUROSCI.1302-13.2014
Costa, A., Torriero, S., & Oliveri, M. (2008). Prefrontal and temporo-parietal involvement in taking others'
perspective: TMS evidence. Behavioural Neurology 19(1-2): 71-4.
Costafreda, S.G., Brammer, M.J., David, A.S., & Fu, C.H. (2008). Predictors of amygdala activation during
the processing of emotional stimuli: A meta-analysis of 385 PET and fMRI studies. Brain Research
Reviews 58(1): 57-70.
Costantini, M. (2005). Neural Systems Underlying Observation of Humanly Impossible Movements: An
fMRI Study. Cereb Cortex 15, 1761-1767.
Coutanche, M. N., Thompson-Schill, S. L., & Schultz, R. T. (2011). Multi-voxel pattern analysis of fMRI
data predicts clinical symptom severity. NeuroImage, 57(1), 113-123. doi:10.1016/j.neuroimage.
2011.04.016
Cox, D. D., & Savoy, R. L. (2003). Functional magnetic resonance imaging (fMRI) 'brain reading':
detecting and classifying distributed patterns of fMRI activity in human visual cortex. NeuroImage,
19(2), 261-270. doi:10.1016/S1053-8119(03)00049-1
Csibra, G., and Gergely, G. (2007). Obsessed with goals': Functions and mechanisms of teleological
interpretation of actions in humans. Acta Psychol 124, 60-78.
Cule, E. (2012). ridge: Ridge Regression with automatic selection of the penalty parameter. R package
version.
Cule, E., & De Iorio, M. (2013). Ridge Regression in Prediction Problems: Automatic Choice of the Ridge
Parameter. Genetic Epidemiology, 37(7), 704-714. doi:10.1002/gepi.21750
Cushman, F. (2008). Crime and punishment: Distinguishing the roles of causal and intentional analyses in
moral judgment. Cognition, 108(2), 353-380. doi:10.1016/j.cognition.2008.03.006
Cushman, F. A. (2008). Crime and punishment: Distinguishing the roles of causal and intentional analyses
in moral judgment. Cognition, 108(2), 353-380. doi:10.1016/j.cognition.2008.03.006
Cushman, F., Young, L.L., & Hauser, M. (2006). The role of conscious reasoning and intuition in moral
judgment testing three principles of harm. Psychological Science, 17(12), 1082-1089.
D'Argembeau, A., Ruby, P., Collette, F., Degueldre, C., Balteau, E., Luxen, A., et al. (2007). Distinct
regions of the medial prefrontal cortex are associated with self-referential processing and perspective
taking. Journal of Cognitive Neuroscience 19(6): 935-944.
Davidson, D. (1963). Actions, reasons, and causes. The Journal of Philosophy, 685-700.
156 of 179
de Gardelle, V., Waszczuk, M., Egner, T., and Summerfield, C. (2012). Concurrent Repetition
Enhancement and Suppression Responses in Extrastriate Visual Cortex. Cereb Cortex. doi: 10.1093/
cercor/bhs211
de Martino, F., Valente, G., Staeren, N., Ashburner, J., Goebel, R., & Formisano, E. (2008). Combining
multivariate voxel selection and support vector machines for mapping and classification of fMRI
spatial patterns. NeuroImage, 43(1), 44-58. doi:10.1016/j.neuroimage.2008.06.037
de Villiers, J. G., Garfield, J., Gernet-Girard, H., Roeper, T., & Speas, M. (2009). Evidentials in Tibetan:
Acquisition, semantics, and cognitive development. New Directions for Child and Adolescent
Development, 2009(125), 29-47. doi:10.1002/cd.248
de Villiers, P.A. (2005). The Role of Language in Theory-of-Mind Development: What Deaf Children Tell
Us A. JW & B. JA, eds. Why language matters for theory of mind.
de Wit, L., Machilsen, B., and Putzeys, T. (2010). Predictive Coding and the Neural Response to
Predictable Stimuli. J Neurosci 30, 8702-8703.
Decety, J., & Lamm, C. (2007). The role of the right temporoparietal junction in social interaction: how
low-level computational processes contribute to meta-cognition. The Neuroscientist : a Review
J o u r n a l B r i n g i n g N e u r o b i o l o g y, N e u r o l o g y a n d P s y c h i a t r y, 1 3 ( 6 ) , 5 8 0 - 5 9 3 . d o i :
10.1177/1073858407304654
Decety, J., Michalska, K. J., & Akitsuki, Y. (2008). Who caused the pain? An fMRI investigation of empathy
and intentionality in children. Neuropsychologia, 46(11), 2607-2614. doi:10.1016/j.neuropsychologia.
2008.05.026
Deen, B., and Saxe, R.R. (2012). Neural correlates of social perception: The posterior superior temporal
sulcus is modulated by action rationality, but not animacy. Proceedings of the 33rd Annual Cognitive
Science Society Conference 276-81.
Dehaene, S., & Cohen, L. (2007). Cultural Recycling of Cortical Maps. Neuron, 56(2), 384-398. doi:
10.1016/j.neuron.2007.10.004
den Ouden, H.E.M., Daunizeau, J., Roiser, J., Friston, K.J., and Stephan, K.E. (2010). Striatal Prediction
Error Modulates Cortical Coupling. J Neurosci 30, 3210-3219.
den Ouden, H.E.M., Friston, K.J., Daw, N.D., McIntosh, A.R., and Stephan, K.E. (2008). A Dual Role for
Prediction Error in Associative Learning. Cereb Cortex 19, 1175-1185.
den Ouden, H.E.M., Kok, P.P., and de Lange, F.P.F. (2012). How prediction errors shape perception,
attention, and motivation. Front Psychol 3, 548-548.
Desimone, R.R., Albright, T.D.T., Gross, C.G.C., and Bruce, C.C. (1984). Stimulus-selective properties of
inferior temporal neurons in the macaque. J Neurosci 4, 2051-2062.
DiCarlo, J. J., & Cox, D. D. (2007). Untangling invariant object recognition. Trends in Cognitive Sciences,
11(8), 333-341. doi:10.1016/j.tics.2007.06.010
DiCarlo, J. J., Zoccolan, D., & Rust, N. C. (2012). How Does the Brain Solve Visual Object Recognition?
Neuron, 73(3), 415-434. doi:10.1016/j.neuron.2012.01.010
Dinstein, I., Thomas, C., Humphreys, K., Minshew, N., Behrmann, M., & Heeger, D. J. (2010). Normal
Movement Selectivity in Autism. Neuron, 66(3), 461-469. doi:10.1016/j.neuron.2010.03.034
Dodell-Feder, D., Koster-Hale, J., Bedny, M., & Saxe, R.R. (2011). fMRI item analysis in a theory of mind
task. NeuroImage, 55(2), 705-712. doi:10.1016/j.neuroimage.2010.12.040
Dodson, C. S., & Johnson, M. K. (1993). Rate of false source attributions depends on how questions are
asked. American Journal of Psychology, 106(4), 541-557.
157 of 179
Döhnel, K., Schuwerk, T., Meinhardt, J., Sodian, B., Hajak, G., & Sommer, M. (2012). Functional activity
of the right temporo-parietal junction and of the medial prefrontal cortex associated with true and false
belief reasoning. NeuroImage, 1-10. doi:10.1016/j.neuroimage.2012.01.073
Dosenbach, N. U. F., Nardos, B., Cohen, A. L., Fair, D. A., Power, J. D., Church, J. A., et al. (2010).
Prediction of Individual Brain Maturity Using fMRI. Science, 329(5997), 1358-1361. doi:10.1126/
science.1194144
Drugowitsch, J.J., and Pouget, A.A. (2012). Probabilistic vs. non-probabilistic approaches to the
neurobiology of perceptual decision-making. Curr Opin Neurobiol 22, 963-969.
Dufour N., Redcay, E, Young, L.L., Mavros, P., Moran, J., Triantafyllou, C., Gabrieli, J., & Saxe, R.R.
(2012). What explains variability in brain regions associated with Theory of Mind in a large sample of
neurotypical adults and adults with ASD? In N. Miyake, D. Peebles, & R. P. Cooper (Eds),
Proceedings of the 34th Annual Conference of the Cognitive Science Society (pp. 312-17). Austin:
Cognitive Science Society.
Dufour, N., Redcay, E., Young, L. L., Mavros, P. L., Moran, J. M., Triantafyllou, C., et al. (2013). Similar
Brain Activation during False Belief Tasks in a Large Sample of Adults with and without Autism. PLoS
ONE, 8(9), e75468. doi:10.1371/journal.pone.0075468
Dunne, S., and O'Doherty, J.P. (2013). Insights from the application of computational neuroimaging to Soc
Neurosci. Curr Opin Neurobiol. 23, 387-392.
Egner, T., Monti, J.M., and Summerfield, C. (2010). Expectation and Surprise Determine Neural
Population Responses in the Ventral Visual Stream. J Neurosci 30, 16601-16608.
Eichenbaum, H. (2004). Hippocampus: cognitive processes and neural representations that underlie
declarative memory. Neuron, 44(1), 109-120.
Ewbank, M.P., Henson, R.N., Rowe, J.B., Stoyanova, R.S., and Calder, A.J. (2013). Different Neural
Mechanisms within Occipitotemporal Cortex Underlie Repetition Suppression across Same and
Different-Size Faces. Cereb Cortex 23, 1073-1084.
Fairhall, S. L., Anzellotti, S., Ubaldi, S., & Caramazza, A. (2013). Person-and place-selective neural
substrates for entity-specific semantic access. Cerebral Cortex, bht039.
Faller, M. (2002). Semantics and Pragmatics of Evidentials in Cuzco Quechua. Doctoral Dissertation,
Stanford University.
Fedorenko, E., & Kanwisher, N. (2009). Neuroimaging of language: Why hasn't a clearer picture
emerged? Language and Linguistics Compass 3(4): 839-65.
Fedorenko, E., Nieto-Casta-Ñn, A., & Kanwisher, N. (2012). Lexical and syntactic representations in the
brain: An fMRI investigation with multi-voxel pattern analyses. Neuropsychologia, 50(4), 499-513. doi:
10.1016/j.neuropsychologia.2011.09.014
Fehr, E., and Camerer, C.F. (2007). Social neuroeconomics: the neural circuitry of social preferences.
Trends Cogn Sci 11, 419-427.
Feldman, H., and Friston, K.J. (2010). Attention, uncertainty, and free-energy. Front Hum Neurosci. 4,
215-215.
Felleman, D.J., and Van Essen, D.C. (1991). Distributed hierarchical processing in the primate Cereb
Cortex. Cereb Cortex 1, 1-47.
Figueras-Costa, B., & Harris, P. (2001). Theory of mind development in deaf children: A nonverbal test of
false-belief understanding. Journal of Deaf Studies and Deaf Education, 6(2), 92.
Fiorillo, C.D., Tobler, P.N., and Schultz, W. (2003). Discrete coding of reward probability and uncertainty
by dopamine neurons. Science 299, 1898-1902.
158 of 179
Fitneva, S. A. (2001). Epistemic marking and reliability judgments: Evidence from Bulgarian. Journal of
Pragmatics, 33, 401-420.
Flesch, R.R. (1948). A new readability yardstick. Journal of Applied Psychology, 32(3), 221-233. doi:
10.1037/h0057532
Fletcher, P., Happe, F., Frith, U., Baker, S.C., Dolan, R.J., Frackowiak, R.S., & Frith, C. D. (1995). Other
minds in the brain: a functional imaging study of 'theory of mind' in story comprehension. Cognition
57(2): 109-28.
Fodor, J. A. (1981). Representations: Philosophical essays on the foundations of cognitive science.
Fodor, J. A. (1987). Modules, frames, fridgeons, sleeping dogs, and the music of the spheres.
Formisano, E., Kim, D. S., Di Salle, F., van de Moortele, P. F., Ugurbil, K., & Goebel, R. (2003). Mirrorsymmetric tonotopic maps in human primary auditory cortex. Neuron, 40(4), 859-869.
Freiwald, W.A., Tsao, D.Y., & Livingstone, M.S. (2009). A face feature space in the macaque temporal
lobe. Nature 12(9): 1187-96.
Friederici, A. D. (2002). Towards a neural basis of auditory sentence processing. Trends in cognitive
sciences, 6(2), 78-84.
Friston, K. (2005). A theory of cortical responses. Philos Trans R Soc Lond B Biol Sci 360, 815-836.
Friston, K. (2009). Modalities, Modes, and Models in Functional Neuroimaging. Science 326, 399-403.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nat Rev Neurosci, 11(2), 127-138.
Friston, K., Holmes, A.P., Worsley, K.J., Poline, J.P., Frith, C.D., & Frackowiak, R.S.J. (1995). Statistical
parametric maps in functional imaging: a general linear approach. Human Brain Mapping 2: 189-210.
Frith, C. D., & Frith, U. (2000). The Physiological Basis of Theory of Mind. In S. L. Baron-Cohen, H. TagerFlusberg, & D. Cohen, Understanding Other Minds: Perspective from Developmental Social
Neuroscience (pp. 335-356). Oxford: Oxford University Press.
Frith, C., & Frith, U. (2012). Social neuroscience. Annual Review of Psychology 63: 287-313.
Furl, N., van Rijsbergen, N. J., Treves, A., Friston, K. J., & Dolan, R. J. (2007). Experience-dependent
coding of facial expression in superior temporal sulcus. Proc Natl Acad Sci U S A, 104(33),
13485-13489.
Gallagher, H. L., Happé, F. G. E., Brunswick, N., Fletcher, P. C., Frith, U., & Frith, C. D. (2000). Reading
the mind in cartoons and stories: an fMRI study of 'theory of mind' in verbal and nonverbal tasks.
Neuropsychologia, 38(1), 11-21.
Gallagher, H., Happé, F., Brunswick, N., Fletcher, P. C., Frith, U., & Frith, C. D. (2000). Reading the mind
in cartoons and stories: an fMRI study of 'theory of mind' in verbal and nonverbal tasks.
Neuropsychologia 38(1): 11-21.
Gallagher, H.L., & Frith, C.D. (2003). Functional imaging of 'theory of mind'. Trends in Cognitive Sciences
7, 77-83.
Gallese, V., & Goldman, A. (1998). Mirror neurons and the simulation theory of mind-reading. Trends in
Cognitive Sciences, 2(12), 493-501.
Gallese, V., & Sinigaglia, C. (2011). What is so special about embodied simulation? Trends in Cognitive
Sciences, 15(11), 512-519. doi:10.1016/j.tics.2011.09.003
Gazzola, V., & Keysers, C. (2009). The Observation and Execution of Actions Share Motor and
Somatosensory Voxels in all Tested Subjects: Single-Subject Analyses of Unsmoothed fMRI Data.
Cerebral Cortex, 19(6), 1239-1255. doi:10.1093/cercor/bhn181
159 of 179
Georgopoulos, A., Schwartz, A., & Kettner, R. (1986). Neuronal population coding of movement direction.
Science 233(4771): 1416-19.
Gergely, G., & Csibra, G. (2003). Teleological reasoning in infancy: The na've theory of rational action.
Trends in Cogn Sci, 7(7), 287-292.
Gilbert, S. J., Meuwese, J. D. I., Towgood, K. J., Frith, C. D., & Burgess, P. W. (2008). Abnormal functional
specialization within medial prefrontal cortex in high-functioning autism: a multi-voxel similarity
analysis. Brain, 132(4), 869-878. doi:10.1093/brain/awn365
Gobbini, M., Koralek, A.C., Bryan, R.E., Montgomery, K.J., & Haxby, J.V. (2007). Two takes on the social
brain: A comparison of theory of mind tasks. Journal of Cognitive Neuroscience 19(11): 1803-14.
Goel, V., Grafman, J., Sadato, N., & Hallett, M. (1995). Modeling other minds. Neuroreport, 6(13),
1741-1746.
Goldman, A. I. (1989). Interpretation Psychologized. Mind & Language, 4(3), 161-185.
Goldman, A. I. (2006). Simulating Minds:The Philosophy, Psychology, and Neuroscience of Mindreading.
Oxford University Press, USA.
Gopnik, A. M., & Meltzoff, A. N. (1997). Words, Thoughts, and Theories. Cambridge, MA: MIT Press.
Gopnik, A. M., & Wellman, H. M. (1992). Why the Child's Theory of Mind Really Is a Theory. Mind &
Language, 7(1-2), 145-171. doi:10.1111/j.1468-0017.1992.tb00202.x
Gopnik, A., & Meltzoff, A.N. (1997). Words, Thoughts, and Theories. Cambridge: MIT Press. 55 Hayward
Street, Cambridge, MA 02142.
Gordon, R. M. (1989). Folk Psychology as Simulation. Mind & Language, 1(2), 158-171. doi:10.1111/j.
1468-0017.1986.tb00324.x
Goris, R. L. T., & Op de Beeck, H. P. (2009). Neural representations that support invariant object
recognition. Frontiers in Computational Neuroscience, 3, 3. doi:10.3389/neuro.10.003.2009
Graesser, A. C., & McNamara, D. S. (2011). Coh-Metrix providing multilevel analyses of text
characteristics. Educational ƒ.
Graf, M., Reitzner, B., Corves, C., Casile, A., Giese, M., and Prinz, W. (2007). Predicting point-light
actions in real-time. NeuroImage 36 Suppl 2, T22-T32.
Grant, C. M. (2005). Moral understanding in children with autism. Autism, 9(3), 317-331. doi:
10.1177/1362361305055418
Greene, J. D., Nystrom, L. E., Engell, A. D., Darley, J. M., & Cohen, J. D. (2004). The Neural Bases of
Cognitive Conflict and Control in Moral Judgment. Neuron, 44(2), 389-400. doi:10.1016/j.neuron.
2004.09.027
Greene, J. D., Sommerville, R. B., Nystrom, L. E., Darley, J. M., & Cohen, J. D. (2001). An fMRI
investigation of emotional engagement in moral judgment. Science, 293(5537), 2105-2108. doi:
10.1126/science.1062872
Gregory, R.L. (1980). Perceptions as hypotheses. Philos Trans R Soc Lond B Biol Sci 290, 181-197.
Grill-Spector, K., Henson, R., & Martin, A. (2006). Repetition and the brain: neural models of stimulusspecific effects. Trends in Cognitive Sciences 10(1): 14-23.
Groen, W. B., Tesink, C., Petersson, K. M., van Berkum, J., van der Gaag, R. J., Hagoort, P., & Buitelaar,
J. K. (2010). Semantic, factual, and social language comprehension in adolescents with autism: an
FMRI study. Cerebral Cortex, 20(8), 1937-1945. doi:10.1093/cercor/bhp264
Grossman, E.D., Battelli, L.L., and Pascual-Leone, A.A. (2005). Repetitive TMS over posterior STS
disrupts perception of biological motion. Vision Res 45, 2847-2853.
160 of 179
Grossman, E.D., Jardine, N. L., & Pyles, J. A. (2010). fMR-adaptation reveals invariant coding of
biological motion on the human STS. Front Hum Neurosci, 4. 15
Gweon, H., Dodell-Feder, D., Bedny, M., & Saxe, R.R. (2012). Theory of Mind Performance in Children
Correlates With Functional Specialization of a Brain Region for Thinking About Thoughts. Child
Development, no-no. doi:10.1111/j.1467-8624.2012.01829.x
Hamilton, D. L., and Sherman, S. J. (1996). Perceiving persons and groups. Psychol Rev, 103(2), 336.
Hamilton, D.L. (1988). Causal attribution viewed from an information-processing perspective. In: D. BarTal & A. W. Kruglanski (Eds), The Social Psychology of Knowledge (pp. 359-85). Cambridge:
Cambridge University Press.
Hamilton, D.L., & Sherman, S.J. (1996). Perceiving persons and groups. Psychological Review 103(2):
336.
Hampton, A. N., & O'Doherty, J. P. (2007). Decoding the neural substrates of reward-related decision
making with functional MRI. Proceedings of the National Academy of Sciences of the United States of
America, 104(4), 1377-1382. doi:10.1073/pnas.0606297104
Hardwig, J. (1991). The role of trust in knowledge. The Journal of Philosophy, 88(12), 693. doi:
10.2307/2027007
Harris, P. (2002a). Checking our sources: the origins of trust in testimony. -Studies in History and
Philosphy of Science, (33), 315-333.
Harris, P. (2002b). What do children learn from testimony? In P. Carruthers, S. Stich, & M. Siegal (Eds.),
The Cognitive Basis of Science. Cambridge University Press.
Harris, P. L. (1992). From Simulation to Folk Psychology: The Case for Development. Mind & Language,
7(1-2), 120-144. doi:10.1111/j.1468-0017.1992.tb00201.x
Harris, P. L., & Koenig, M. A. (2006). Trust in testimony: how children learn about science and religion.
Child Development, 77(3), 505-524. doi:10.1111/j.1467-8624.2006.00886.x
Harrison, S. A., & Tong, F. (2009). Decoding reveals the contents of visual working memory in early visual
areas. Nature, 458(7238), 632-635.
Hassabis, D., Spreng, R. N., Rusu, A. A., Robbins, C. A., Mar, R. A., & Schacter, D. L. (2013). Imagine All
the People: How the Brain Creates and Uses Personality Models to Predict Behavior. Cerebral Cortex
(New York, N.Y. : 1991), 24(8), bht042-1987. doi:10.1093/cercor/bht042
Hastie, T. J., Tibshirani, R. J., & Friedman, J. H. (2011). The elements of statistical learning: data mining,
inference, and prediction.
Haxby, J. V. (2001). Distributed and Overlapping Representations of Faces and Objects in Ventral
Temporal Cortex. Science, 293(5539), 2425-2430. doi:10.1126/science.1063736
Hayasaka, S., & Nichols, T. E. (2004). Combining voxel intensity and cluster extent with permutation test
framework. NeuroImage, 23(1), 54-63. doi:10.1016/j.neuroimage.2004.04.035
Hayashi, A., Abe, N., Ueno, A., Shigemune, Y., Mori, E., Tashiro, M., & Fujii, T. (2010). Neural correlates of
forgiveness for moral transgressions involving deception. Brain Research 1332: 90-9.
Hayden, B.Y., Heilbronner, S.R., Pearson, J.M., and Platt, M.L. (2011). Surprise Signals in Anterior
Cingulate Cortex: Neuronal Encoding of Unsigned Reward Prediction Errors Driving Adjustment in
Behavior. J Neurosci. 31, 4178-4187.
Haynes, J.-D., & Rees, G. (2006). Decoding mental states from brain activity in humans. Nature Reviews
Neuroscience, 7(7), 523-534. doi:10.1038/nrn1931
He, D., Kersten, D., and Fang, F. (2012). Opposite Modulation of High- and Low-Level Visual Aftereffects
by Perceptual Grouping. Curr Biol 22, 1040-1045.
161 of 179
Hearst, M. A., Dumais, S. T., Osman, E., Platt, J., & Scholkopf, B. (1998). Support vector machines.
Intelligent Systems and Their Applications, IEEE, 13(4), 18-28. doi:10.1109/5254.708428
Heider, F., & Simmel, M. (1944). An experimental study of apparent behavior. American Journal of
Psychology 57(2): 243-59.
Hein, G., & Knight, R.T. (2008). Superior temporal sulcus-it's my area: or is it? Journal of Cognitive
Neuroscience 20(12): 2125-36.
Herrmann, K., Montaser-Kouhsari, L., Carrasco, M., and Heeger, D.J. (2010). When size matters:
attention affects performance by contrast or response gain. Nature 13, 1554-1559.
Hesselmann, G.G., Sadaghiani, S.S., Friston, K.J.K., and Kleinschmidt, A.A. (2010). Predictive coding or
evidence accumulation? False inference and neuronal fluctuations. PLoS ONE 5, e9926-e9926.
Heyes, C. (2014). False belief in infancy: a fresh look. Developmental Science, n/a-n/a. doi:10.1111/desc.
12148
Higgins, E.T., & Bargh, J.A. (1987). Social cognition and social perception. Annual Review of Psychology
38(1): 369-425.
Hillyard, S.A., Vogel, E.K., and Luck, S.J. (1998). Sensory gain control (amplification) as a mechanism of
selective attention: electrophysiological and neuroimaging evidence. Philos Trans R Soc Lond B Biol
Sci 353, 1257-1270.
Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems.
Technometrics, 12(1), 55-67.
Horgan, T. (1992). Review Essay: From Cognitive Science to Folk Psychology: Computation, Mental
Representation, and Belief. Philosophy and Phenomenological Research, 449-484.
Hung, C. P., Kreiman, G., Poggio, T., & DiCarlo, J. J. (2005). Fast Readout of Object Identity from
Macaque Inferior Temporal Cortex. Science, 310(5749), 863-866. doi:10.1126/science.1117593
Immordino-Yang, M. H., McColl, A., Damasio, H., & Damasio, A. (2009). Neural correlates of admiration
and compassion. Proceedings of the National Academy of Sciences of the United States of America,
106(19), 8021-8026. doi:10.1073/pnas.0810363106
Ishai, A., Pessoa, L., Bikle, P. C., & Ungerleider, L. G. (2004). Repetition suppression of faces is
modulated by emotion. Proc Natl Acad Sci U S A, 101(26), 9827-9832.
Izuma, K., Saito, D. N., & Sadato, N. (2008). Processing of Social and Monetary Rewards in the Human
Striatum. Neuron, 58(2), 284-294. doi:10.1016/j.neuron.2008.03.020
Izuma, K., Saito, D.N., and Sadato, N. (2008). Processing of Social and Monetary Rewards in the Human
Striatum. Neuron 58, 284-294.
Jacques, P., Conway, M.A., Lowder, M.W., & Cabeza, R. (2011). Watching my mind unfold versus yours:
An fMRI study using a novel camera technology to examine neural differences in self-projection of
self vs. other perspectives. Journal of Cognitive Neuroscience 23(6): 1275-84
Jastorff, J., Clavagnier, S., Gergely, G., and Orban, G.A. (2011). Neural Mechanisms of Understanding
Rational Actions: Middle Temporal Gyrus Activation by Contextual Violation. Cereb Cortex 21,
318-329.
Jaswal, V. K., & Neely, L. A. (2006). Adults don't always know best: preschoolers use past reliability over
age when learning new words. Psychological Science, 17(9), 757-758. doi:10.1111/j.
1467-9280.2006.01778.x
Jehee, J. F., & Ballard, D. H. (2009). Predictive feedback can account for biphasic responses in the lateral
geniculate nucleus. PLoS computational biology, 5(5), e1000373.
162 of 179
Jellema, T., and Perrett, D.I. (2003a). Cells in monkey STS responsive to articulated body motions and
consequent static posture: a case of implied motion? Neuropsychologia 41, 1728-1737.
Jellema, T., and Perrett, D.I. (2003b). Perceptual history influences neural responses to face and body
postures. J Cogn Neurosci 15, 961-971.
Jellema, T., Baker, C.I., Wicker, B., and Perrett, D.I. (2000). Neural Representation for the Perception of
the Intentionality of Actions. Brain Cogn 44, 23-23.
Jenkins, A. C., & Mitchell, J. P. (2010). Mentalizing under uncertainty: dissociated neural responses to
ambiguous and unambiguous mental state inferences. Cerebral Cortex, 20(2), 404-410. doi:10.1093/
cercor/bhp109
Jenkins, A., Macrae, C., & Mitchell, J. (2008). Repetition suppression of ventromedial prefrontal activity
during judgments of self and others. Proceedings of the National Academy of Sciences 105(11):
4507.
Johanson, L., & Utas, B. (2000). Evidentials. Walter de Gruyter.
Jones, R.M., Somerville, L.H., Li, J., Ruberry, E.J., Libby, V., Glover, G., Voss, H.U., Ballon, D.J., and
Casey, B.J. (2011). Behavioral and Neural Properties of Social Reinforcement Learning. J Neurosci
31, 13039-13045.
Joseph, R. M., Tager-Flusberg, H., & Lord, C. (2002). Cognitive profiles and social-communicative
functioning in children with autism spectrum disorder. Journal of Child Psychology and Psychiatry,
43(6), 807-821.
Kable, J. W., & Chatterjee, A. (2006). Specificity of action representations in the lateral occipitotemporal
cortex. J Cogn Neurosci, 18(9), 1498-1517.
Kalbe, E., Schlegel, M., Sack, A.T., Nowak, D.A., Dafotakis, M., Bangard, C., et al. (2010). Dissociating
cognitive from affective theory of mind: a TMS study. Cortex, 46(6): 769-80.
Kamitani, Y., & Tong, F. (2005). Decoding the visual and subjective contents of the human brain. Nature
Neuroscience, 8(5), 679-685. doi:10.1038/nn1444
Kamitani, Y., & Tong, F. (2006). Decoding Seen and Attended Motion Directions from Activity in the
Human Visual Cortex. Current Biology, 16(11), 1096-1102. doi:10.1016/j.cub.2006.04.003
Kanwisher, N. (2010). Functional specificity in the human brain: a window into the functional architecture
of the mind. Proceedings of the National Academy of Sciences 107(25): 11163-70.
Kanwisher, N., McDermott, J., and Chun, M. (1997). The fusiform face area: a module in human
extrastriate cortex specialized for face perception. J Neurosci 17, 4302-4311.
Kassam, K. S., Markey, A. R., Cherkassky, V. L., Loewenstein, G., & Just, M. A. (2013). Identifying
Emotions on the Basis of Neural Activation. PLoS ONE, 8(6), e66032. doi:10.1371/journal.pone.
0066032
Keller, G.B., Bonhoeffer, T., and Hübener, M. (2012). Sensorimotor Mismatch Signals in Primary Visual
Cortex of the Behaving Mouse. Neuron 74, 809-815.
Kemp, C., & Tenenbaum, J. B. (2008). The discovery of structural form.Proceedings of the National
Academy of Sciences, 105(31), 10687-10692.
Kennedy, D. P., & Courchesne, E. (2008). Functional abnormalities of the default network during self- and
other-reflection in autism. Social Cognitive and Affective Neuroscience, 3(2), 177-190. doi:10.1093/
scan/nsn011
Keysers, C., & Gazzola, V. (2007). Integrating simulation and theory of mind: from self to social cognition.
Trends in Cognitive Sciences, 11(5), 194-196. doi:10.1016/j.tics.2007.02.002
163 of 179
Kilner, J. M., & Frith, C. D. (2007). Action observation: inferring intentions without mirror neurons. Current
Biology, 18(1), R32-3. doi:10.1016/j.cub.2007.11.008
Kilner, J. M., Paulignan, Y., & Blakemore, S. J. (2003). An Interference Effect of Observed Biological
Movement on Action. Current Biology, 13(6), 522-525. doi:10.1016/S0960-9822(03)00165-9
Kincaid, J. P., Fishburne, R. P., Jr, Rogers, R. L., & Chissom, B. S. (1975). Derivation of new readability
formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted
personnel.
Knight, R.T. (2007). Neural networks debunk phrenology. Science 316(5831): 1578-9.
Kobayashi, C., Glover, G. H., & Temple, E. (2007). Children's and adults' neural bases of verbal and
nonverbal 'theory of mind'. Neuropsychologia, 45(7), 1522-1532. doi:10.1016/j.neuropsychologia.
2006.11.017
Kobayashi, C., Glover, G.H., & Temple, E. (2007). Children's and adults'neural bases of verbal and
nonverbal 'theory of mind'. Neuropsychologia 45: 1522-32.
Koenig, M. A., & Harris, P. L. (2005a). Preschoolers mistrust ignorant and inaccurate speakers. Child
Development, 76(6), 1261-1277. doi:10.1111/j.1467-8624.2005.00849.x
Koenig, M. A., & Harris, P. L. (2005b). The role of social cognition in early trust. Trends in Cognitive
Sciences, 9(10), 457-459. doi:10.1016/j.tics.2005.08.006
Koenig, M. A., Clément, F. & Harris, P. L. (2004). Trust in testimony: children's use of true and false
statements. Psychological Science, 15(10), 694-698. doi:10.1111/j.0956-7976.2004.00742.x
Koenigs, M., Young, L.L., Adolphs, R., Tranel, D., Cushman, F., Hauser, M., et al. (2007). Damage to the
prefrontal cortex increases utilitarian moral judgements. Nature 446(7138): 908-11.
Kok, P. P., Jehee, J. F. M. J., & de Lange, F. P. F. (2012). Less is more: expectation sharpens
representations in the primary visual cortex. Neuron, 75(2), 265-270. doi:10.1016/j.neuron.
2012.04.034
Kok, P.P., Jehee, J.F.M.J., and de Lange, F.P.F. (2012a). Expectation Sharpens Representations in the
Primary Visual Cortex. Neuron 75, 265-270.
Kok, P.P., Rahnev, D.D., Jehee, J.F.M.J., Lau, H.C.H., and de Lange, F.P.F. (2012b). Attention reverses
the effect of prediction in silencing sensory signals. Cereb Cortex 22, 2197-2206.
Konkle, T., & Caramazza, A. (2013). Tripartite Organization of the Ventral Stream by Animacy and Object
Size. The Journal of Neuroscience : the Official Journal of the Society for Neuroscience, 33(25),
10235-10242. doi:10.1523/JNEUROSCI.0983-13.2013
Konkle, T., & Oliva, A. (2012). A real-world size organization of object responses in occipto-temporal
cortex. Neuron 74(6): 1114-24.
Kornblith, H. (1983). Justified belief and epistemically responsible action. The Philosophical Review.
Koshino, H., Kana, R. K., Keller, T. A., Cherkassky, V. L., Minshew, N. J., & Just, M. A. (2007). fMRI
Investigation of Working Memory for Faces in Autism: Visual Coding and Underconnectivity with
Frontal Areas. Cerebral Cortex, 18(2), 289-300. doi:10.1093/cercor/bhm054
Koster-Hale, J., & Saxe, R.R. (2011). Theory of Mind brain regions are sensitive to the content, not the
structural complexity, of belief attributions. In: L. Carlson, C. Hoelscher, & T. F. Shipley (Eds),
Proceedings of the 33rd Annual Cognitive Science Society Conference, pp. 3356-61.
Koster-Hale, J., & Saxe, R.R. (2013). Functional neuroimaging of theory of mind. In S. Baron-Cohen, H.
Tager-Flusberg, & M. V. Lombardo, Understanding Other Minds Perspectives from developmental
social neuroscience (pp. 132-163). Oxford University Press.
164 of 179
Koster-Hale, J., & Saxe, R.R. (2013). Theory of mind: a neural prediction problem. Neuron, 79(5), 836–
848. doi:10.1016/j.neuron.2013.08.020
Koster-Hale, J., Bedny, M., & Saxe, R.R. (2014). Thinking about seeing: Perceptual sources of knowledge
are encoded in the theory of mind brain regions of sighted and blind adults. Cognition, 133(1), 65-78.
doi:10.1016/j.cognition.2014.04.006
Koster-Hale, J., Saxe, R.R., Dungan, J., & Young, L.L. (2013). Decoding moral judgments from neural
representations of intentions. Proceedings of the National Academy of Sciences of the United States
of America, 110(14), 5648-5653. doi:10.1073/pnas.1207992110
Kourtzi, Z. & Kanwisher, N. (2001). Representation of perceived object shape by the human lateral
occipital complex. Science 293(5534): 1506.
Kourtzi, Z., and Kanwisher, N. (2000). Activation in human MT/MST by static images with implied motion.
J Cogn Neurosci 12, 48-55.
Krall, S. C., Rottschy, C., Oberwelland, E., Bzdok, D., Fox, P. T., Eickhoff, S. B., et al. (2014). The role of
the right temporoparietal junction in attention and social interaction as revealed by ALE metaanalysis. Brain Structure and Function. doi:10.1007/s00429-014-0803-z
Kriegeskorte, N. (2008). Representational similarity analysis - connecting the branches of systems
neuroscience. Frontiers in Systems Neuroscience. doi:10.3389/neuro.06.004.2008
Kriegeskorte, N., & Bandettini, P. (2007). Analyzing for information, not activation, to exploit highresolution fMRI. NeuroImage, 38(4), 649-662. doi:10.1016/j.neuroimage.2007.02.022
Kriegeskorte, N., & Kievit, R. A. (2013). Representational geometry: integrating cognition, computation,
and the brain. Trends in Cognitive Sciences, 17(8), 401-412. doi:10.1016/j.tics.2013.06.007
Kriegeskorte, N., Bandettini, P., & Bandettini, P. (2007). Analyzing for information, not activation, to exploit
high-resolution fMRI. NeuroImage, 38(4), 649-662. doi:10.1016/j.neuroimage.2007.02.022
Kriegeskorte, N., Goebel, R., & Bandettini, P.A. (2006). Information-based functional brain mapping.
Proceedings of the National Academy of Sciences USA 103(10), 3863-8.
Kriegeskorte, N., Mur, M., Ruff, D. A., Kiani, R., Bodurka, J., Esteky, H., ... & Bandettini, P. A. (2008).
Matching categorical object representations in inferior temporal cortex of man and monkey. Neuron,
60(6), 1126-1141.
Krienen, F.M., Tu, P.C., & Buckner, R.L. (2010). Clan mentality: Evidence that the medial prefrontal cortex
responds to close others. Journal of Neuroscience 30(41): 13906-15.
Krömer, N., Schöfer, J., & Boulesteix, A.-L. (2009). Regularized estimation of large-scale gene association
networks using graphical Gaussian models. BMC Bioinformatics, 10(1), 384. doi:
10.1186/1471-2105-10-384
Krueger, J., and Clement, R. W. (1994). The truly false consensus effect: An ineradicable and edocentric
bias in social perception. J Pers Soc Psychol, 67, 596 - 610.
Kubit, B., & Jack, A. I. (2013). Rethinking the role of the rTPJ in attention and social cognition in light of
the opposing domains hypothesis: findings from an ALE-based meta-analysis and resting-state
functional connectivity. Frontiers in Human Neuroscience, 7, 323. doi:10.3389/fnhum.2013.00323
Lamm, C., Batson, C.D., & Decety, J. (2007). The neural substrate of human empathy: effects of
perspective-taking and cognitive appraisal. Journal of Cognitive Neuroscience 19(1): 42-58.
Landau, B., & Gleitman, L. R. (1985). Language and Experience. Harvard University Press.
Lane, R.D., Chua, P.M., and Dolan, R.J. (1999). Common effects of emotional valence, arousal and
attention on neural activation during visual processing of pictures. Neuropsychologia 37, 989-997.
165 of 179
Lawrence, E.J., Shaw, P., Baker, D., Baron-Cohen, S., & David, A.S. (2004). Measuring empathy:
reliability and validity of the empathy quotient. Psychological Medicine, 34(5): 911-24.
Lazebnik, S., Schmid, C., & Ponce, J. (2006). Beyond bags of features: Spatial pyramid matching for
recognizing natural scene categories. In Computer Vision and Pattern Recognition, 2006 IEEE
Computer Society Conference on (Vol. 2, pp. 2169-2178). IEEE.
Lenci, A., Baroni, M., Cazzolli, G., & Marotta, G. (2013). BLIND: a set of semantic feature norms from the
congenitally blind. Behavior Research Methods, 1-16. doi:10.3758/s13428-013-0323-4
Leshinskaya, A., & Caramazza, A. (2014). Nonmotor Aspects of Action Concepts. Journal of Cognitive
Neuroscience, 15, 1-17. doi:10.1162/jocn-a-00679
Leslie, A., Mallon, R., & DiCorcia, J. (2006). Transgressors, victims, and cry babies: Is basic moral
judgment spared in autism? Social Neuroscience, 1(3-4), 270-283.
Leslie, A.M., & Frith, U. (1990). Prospects for a cognitive neuropsychology of autism: Hobson's choice.
Psychological Review 97(1): 122-31.
Lin, A., Adolphs, R., and Rangel, A. (2012). Social and monetary reward learning engage overlapping
neural substrates. Soc Cogn Affect Neurosci 7, 274-281.
Lindsay, D. S., & Johnson, M. K. (1989). The eyewitness suggestibility effect and memory for source.
Memory and Cognition, 17(3), 349-358. doi:10.1016/j.jml.2003.08.007
Liu, D., Sabbagh, M., A., Gehring, W.J., & Wellman, H.M. (2004). Decoupling beliefs from reality in the
brain: an ERP study of theory of mind. Neuroreport 15(6): 991-5.
Logothetis, N. K., & Sheinberg, D. L. (1996). Visual object recognition. Ann Rev Neurosci, 19(1), 577-621.
Logothetis, N.K. (2008). What we can do and what we cannot do with fMRI. Nature, 453(7197): 869-78.
Lombardo, M. V., Chakrabarti, B., Bullmore, E. T., Baron-Cohen, S., & Consortium, M. A. (2011).
Specialization of right temporo-parietal junction for mentalizing and its relation to social impairments
in autism. NeuroImage, 56(3), 1832-1838. doi:10.1016/j.neuroimage.2011.02.067
Lombardo, M. V., Chakrabarti, B., Bullmore, E. T., Wheelwright, S. J., Sadek, S. A., Suckling, J., & BaronCohen, S. L. (2010). Shared neural circuits for mentalizing about the self and others. Journal of
Cognitive Neuroscience, 22(7), 1623-1635. doi:10.1162/jocn.2009.21287
Lombardo, M.V., Chakrabarti, B., Bullmore, E.T., Wheelwright, S.J., Sadek, S.A., Suckling, J., et al.
(2010). Shared neural circuits for mentalizing about the self and others. Journal of Cognitive
Neuroscience, 22(7): 1623-35.
Lord, C., Risi, S., Lambrecht, L., Cook, E. H., Leventhal, B. L., DiLavore, P.C., et al. (2000). The Autism
Diagnostic Observation Schedule„Generic: A standard measure of social and communication deficits
associated with the spectrum of autism. Journal of Autism and Developmental Disorders, 30(3),
205-223. doi:10.1023/A:1005592401947
Lucas, A. J., Lewis, C., Pala, F. C., Wong, K., & Berridge, D. (2013). Social-cognitive processes in
preschoolers' selective trust: Three cultures compared. Developmental Psychology, 49(3), 579-590.
doi:10.1037/a0029864
Ma, N., Baetens, K., Vandekerckhove, M., Kestemont, J., Fias, W., & Van Overwalle, F. (2013a). Traits
are represented in the medial prefrontal cortex: an fMRI adaptation study. Social Cognitive and
Affective Neuroscience. doi:10.1093/scan/nst098
Ma, N., Baetens, K., Vandekerckhove, M., Van der Cruyssen, L., & Van Overwalle, F. (2013b).
Dissociation of a trait and a valence representation in the mPFC. Social Cognitive and Affective
Neuroscience. doi:10.1093/scan/nst143
Ma, N., Vandekerckhove, M., Baetens, K., Van Overwalle, F., Seurinck, R., and Fias, W. (2012).
Inconsistencies in spontaneous and intentional trait inferences. Soc Cogn Affect Neurosci 7, 937-950.
166 of 179
Mahon, B. Z., & Caramazza, A. (2010). Judging semantic similarity: an event-related fMRI study with
auditory word stimuli. Nsc, 169(1), 279-286. doi:10.1016/j.neuroscience.2010.04.029
Malle, B. F. (1999). How people explain behavior: A new theoretical framework. Pers Soc Psychol Rev
3(1), 23-48.
Manera, V., Becchio, C., Ben Schouten, Bara, B.G., and Verfaillie, K. (2011). Communicative interactions
improve visual detection of biological motion. PLoS ONE 6, e14594-e14594.
Marmor, G. S. (n.d.). Age at onset of blindness and the development of the semantics of color names.
Journal of Experimental Child Psychology, 25(2), 267-278.
Mars, R. B., Sallet, J., Schüffelgen, U., Jbabdi, S., Toni, I., & Rushworth, M. F. S. (2012). Connectivitybased subdivisions of the human right 'temporoparietal junction area': evidence for different areas
participating in different cortical networks. Cerebral Cortex (New York, N.Y. : 1991), 22(8), 1894-1903.
doi:10.1093/cercor/bhr268
Mars, R.B., Sallet, J., Neubert, F.-X., and Rushworth, M.F.S. (2013). Connectivity profiles reveal the
relationship between brain areas for social cognition in human and monkey temporoparietal cortex.
Proc Natl Acad Sci U S A. doi:10.1073/pnas.1302956110
Martinez-Trujillo, J.C.J., and Treue, S.S. (2004). Feature-Based Attention Increases the Selectivity of
Population Responses in Primate Visual Cortex. Curr Biol 14, 8-8.
Mason, R. A., & Just, M. A. (2011). Differentiable cortical networks for inferences concerning people's
intentions versus physical causality. Human Brain Mapping, 32(2), 313-329. doi:10.1002/hbm.21021
Mason, R. A., Williams, D. L., Kana, R. K., Minshew, N., & Just, M. A. (2008). Theory of Mind disruption
and recruitment of the right hemisphere during narrative comprehension in autism. Neuropsychologia,
46(1), 269-280. doi:10.1016/j.neuropsychologia.2007.07.018
Mason, R.A., & Just, M.A. (2010). Differentiable cortical networks for inferences concerning people's
intentions vs. physical causality. Human Brain Mapping, 32(2): 313-29.
Maunsell, J. H., and Newsome, W. T. (1987). Visual processing in monkey extrastriate cortex. Ann Rev
Neurosci 10(1), 363-401.
McAlpine, L., & Moore, C. D. (1995). The development of social understanding in children with visual
impairments. Journal of Visual Impairment and Blindness, (89), 349-358.
McNamara, D. S., Graesser, A. C., McCarthy, P. M., & Cai, Z. (2014). Automated Evaluation of Text and
Discourse with Coh-Metrix. Cambridge University Press.
Meltzoff, A. N. (2004). The case for a developmental cognitive science: Theories of people and things. In
G. Brenmner & A. Slater (Eds.), Theories of infant development (pp. 145-173). Oxford: Blackwell.
Meltzoff, A. N. (2007). 'Like me': a foundation for social cognition. Developmental Science, 10(1),
126-134. doi:10.1111/j.1467-7687.2007.00574.x
Meltzoff, A. N., & Brooks, R. (2007). Eyes wide shut: The importance of eyes in infant gaze following and
understanding of other minds. In R. Flom, K. Lee, & D. Muir (Eds.), Gaze Following: its development
and signficance (pp. 217-241). Eyes wide shut: The importance of eyes in infant gaze following and
understanding of other minds.: Eyes wide shut: The importance of eyes in infant gaze following and
understanding of other minds.
Mende-Siedlecki, P., Cai, Y., and Todorov, A. (2012). The neural dynamics of updating person
impressions. Soc Cogn Affect Neurosci. doi: 10.1093/scan/nss040
Meristo, M., Falkman, K. W., Hjelmquist, E., Tedoldi, M., Surian, L., & Siegal, M. (2007). Language access
and theory of mind reasoning: Evidence from deaf children in bilingual and oralist environments.
Developmental Psychology, 43(5), 1156-1169. doi:10.1037/0012-1649.43.5.1156
167 of 179
Meyer, M., and Sauerland, U. (2009). A pragmatic constraint on ambiguity detection. Nat Lang Linguist
Theory 27, 139-150.
Meyer, T., and Olson, C. R. (2011). Statistical learning of visual transitions in monkey inferotemporal
cortex. Proc Natl Acad Sci U S A, 108(48), 19401-19406.
Miene, P., Bordiga, E., & Park, R. (1993). The Evaluation of Hearsay Evidence: A Social Psychological
Approach. In N. J. Castellan (Ed.), Individual and Group Decision Making. Psychology Press.
Miene, P., Park, R. C., & Bordiga, E. (1992). Juror Decision Making and the Evaluation of Hearsay
Evidence. 76 Minn . L. Rev, (683).
Millar, S. (1976). Spatial representation by blind and sighted children. Journal of Experimental Child
Psychology, 21(3), 460-479.
Miller, E.K., & Cohen, J.D. (2001). An integrative theory of prefrontal cortex function. Annual Review of
Neuroscience, 24: 167-202.
Minter, M., Hobson, R. P., & B, M. (1998). Congenital visual impairment and 'theory of mind.' British
Journal of Developmental Psychology, (16), 183-196.
Mitchell, J., & Banaji, M. (2005). The link between social cognition and self-referential thought in the
medial prefrontal cortex. Journal of Cognitive Neuroscience.
Mitchell, J., Macrae, C.N., & Banaji, M.R. (2006). Dissociable medial prefrontal contributions to judgments
of similar and dissimilar others. Neuron, 50(4): 655-63.
Mitchell, J.P. (2008). Activity in right temporo-parietal junction is not selective for theory-of-mind. Cerebral
Cortex, 18(2): 262-71.
Mitchell, J.P., Banaji, M.R., & MacRae, C.N. (2005). The link between social cognition and self-referential
thought in the medial prefrontal cortex. Journal of Cognitive Neuroscience 17(8): 1306-15.
Mitchell, J.P., Macrae, C.N., and Banaji, M.R. (2006). Dissociable Medial Prefrontal Contributions to
Judgments of Similar and Dissimilar Others. Neuron 50, 655-663.
Mitchell, T. M., Hutchinson, R., Niculescu, R. S., Pereira, F., Wang, X., Just, M., & Newman, S. (2004).
Learning to Decode Cognitive States from Brain Images. Machine Learning, 57(1/2), 145-175. doi:
10.1023/B:MACH.0000035475.85309.1b
Moeller, M., & Schick, B. (2006). Relations between maternal input and theory of mind understanding in
deaf children. Child Development, 77(3), 751-766.
Moran, J. M., Young, L.L., Saxe, R.R., Lee, S. M., O'Young, D., Mavros, P. L., & Gabrieli, J. D. (2011).
Impaired theory of mind for moral judgment in high-functioning autism. Proceedings of the National
Academy of Sciences of the United States of America, 108(7), 2688-2692. doi:10.1073/pnas.
1011734108
Moran, J.M., Lee, S.M., & Gabrieli, J.D.E. (2011a). Dissociable neural systems supporting knowledge
about human character and appearance in ourselves and others. Journal of Cognitive Neuroscience,
23(9): 2222-30.
Moran, J.M., Young, L.L., Saxe, R.R., Lee, S.M., O'Young, D., Mavros, P.L., & Gabrieli, J.D. (2011b).
Impaired theory of mind for moral judgment in high-functioning autism. Proceedings of the National
Academy of Sciences, 108(7): 2688-92.
Moriguchi, Y., Ohnishi, T., Lane, R. D., Maeda, M., Mori, T., Nemoto, K., et al. (2006). Impaired selfawareness and theory of mind: An fMRI study of mentalizing in alexithymia. NeuroImage, 32(3),
1472-1482. doi:10.1016/j.neuroimage.2006.04.186
Morton, J. (1994). Cognitive perspectives on memory recovery. Applied Cognitive Psychology, 8, 389-398.
168 of 179
Müller, N.G., Strumpf, H., Scholz, M., Baier, B., and Melloni, L. (2013). Repetition suppression versus
enhancement—it's quantity that matters. Cereb Cortex 23, 315-322.
Mumford, D. (1992). On the computational architecture of the neocortex. II. The role of cortico-cortical
loops. Biol Cybern 66, 241-251.
Murray, S.O., Kersten, D., Olshausen, B.A., Schrater, P., and Woods, D.L. (2002). Shape perception
reduces activity in human primary visual cortex. Proc Natl Acad Sci U S A 99, 15164-15169.
Nakahara, H., Itoh, H., Kawagoe, R., Takikawa, Y., & Hikosaka, O. (2004). Dopamine neurons can
represent context-dependent prediction error. Neuron,41(2), 269-280.
Naselaris, T., Kay, K. N., Nishimoto, S., & Gallant, J. L. (2011). Encoding and decoding in fMRI.
NeuroImage, 56(2), 400-410. doi:10.1016/j.neuroimage.2010.07.073
Neri, P.P., Luu, J.Y.J., and Levi, D.M.D. (2006). Meaningful interactions can enhance visual discrimination
of human agents. Nat Neurosci 9, 1186-1192.
Nichols, S., & Ulatowski, J. (2007). Intuitions and Individual Differences: The Knobe Effect Revisited. Mind
& Language, 22(4), 346-365. doi:10.1111/j.1468-0017.2007.00312.x
Nichols, S., Stich, S., Leslie, A., & Klein, D. (1996). Varieties of Off-Line Simulation. Theories of Theories
of Mind.
Nichols, T. (2002). Nonparametric permutation tests for functional neuroimaging: a primer with examples.
Human Brain Mapping.
Nieminen-von Wendt, T., Mets honkala, L., Kulom ki, T., Alto, S., Autti, T., Vanhala, R., & Wendt, von, L.
(2003). Changes in cerebral blood flow in Asperger syndrome during theory of mind tasks presented
by the auditory route. European Child & Adolescent Psychiatry, 12(4), 178-189. doi:10.1007/
s00787-003-0337-z
Norman, K., Polyn, S., Detre, G., & Haxby, J. (2006). Beyond mind-reading: multi-voxel pattern analysis of
fMRI data. Trends in Cognitive Sciences, 10(9), 424-430.
Norman-Haignere S, McDermott JH, Kanwisher N (February 2014). Data-driven clustering of cortical
responses to natural sounds reveals organization for pitch and speech. Poster presented at the
Computational and Systems Neuroscience meeting (COSYNE), Salt Lake City, UT.
O'Doherty, J., Kringelbach, M. L., Rolls, E. T., Hornak, J., & Andrews, C. (2001). Abstract reward and
punishment representations in the human orbitofrontal cortex. Nature neuroscience, 4(1), 95-102.
O'Neill, D. K., & Chong, S. C. (2001). Preschool children's difficulty understanding the types of information
obtained through the five senses. Child Development, 72(3), 803-815. doi:10.2307/1132456
O'Neill, D. K., & Gopnik, A. (1991). Young children's ability to identify the sources of their beliefs.
Developmental Psychology, 27(3), 390-397.
Olson, G. (2003). Reconsidering Unreliability: Fallible and Untrustworthy Narrators. Narrative, 11(1),
93-109. doi:10.1353/nar.2003.0001
Onishi, K., & Baillargeon, R. (2005). Do 15-month-old infants understand false beliefs? Science,
308(5719): 255.
Onishi, K., Baillargeon, R., & Baillargeon, R. (2005). Do 15-month-old infants understand false beliefs?
Science, 308(5719), 255.
Ozonoff, S., Pennington, B. F., & Rogers, S. J. (1991). Executive function deficits in high‐functioning
autistic individuals: relationship to theory of mind.Journal of child Psychology and Psychiatry, 32(7),
1081-1105.
169 of 179
Papafragou, A., & Li, P. (2001). Evidential morphology and theory of mind. Presented at the Proceedings
from the 26th Annual Boston University Conference on Language Development, Cascadilla Press,
Somerville, MA.
Park, R. C. (1987). A Subject Matter Approach to Hearsay Reform. 86 Mich. L. Rev. (51).
Patel, D., Fleming, S.M., and Kilner, J.M. (2012). Inferring subjective states through the observation of
actions. Philos Trans R Soc Lond B Biol Sci 279, 4853-4860.
Pavlova, M., Staudt, M., Sokolov, A., Birbaumer, N., & Krögeloh-Mann, I. (2003). Perception and
production of biological movement in patients with early periventricular brain lesions. Brain, 126(3),
692-701.
Peelen, M. V., Atkinson, A. P., & Vuilleumier, P. (2010). Supramodal Representations of Perceived
Emotions in the Human Brain. Journal of Neuroscience, 30(30), 10127-10134. doi:10.1523/
JNEUROSCI.2161-10.2010
Peelen, M. V., Wiggett, A. J., & Downing, P. E. (2006). Patterns of fMRI Activity Dissociate Overlapping
Functional Brain Areas that Respond to Biological Motion. Neuron, 49(6), 815-822. doi:10.1016/
j.neuron.2006.02.004
Peelen, M.V.M., and Kastner, S.S. (2011). A neural basis for real-world visual search in human
occipitotemporal cortex. Proc Natl Acad Sci U S A 108, 12125-12130.
Pelphrey, K.A., & Morris, J.P. (2006). Brain mechanisms for interpreting the actions of others from
biological-motion cues. Current Directions in Psychological Science, 15(3): 136.
Pelphrey, K.A., and Vander Wyk, B.C. (2011). Functional and neural mechanisms for eye gaze
processing. In OUP Handbook of Face Perception, A Calder, G Rhodes, M Johnson, J Haxby, eds.
(Oxford University Press), pp:591-604.
Pelphrey, K.A., Mitchell, T.V., McKeown, M.J., Goldstein, J., Allison, T., & McCarthy, G. (2003). Brain
activity evoked by the perception of human walking: Controlling for meaningful coherent motion.
Journal of Neuroscience, 23, 6819-25.
Pelphrey, K.A., Morris, J.P., and McCarthy, G. (2004). Grasping the intentions of others: the perceived
intentionality of an action influences activity in the superior temporal sulcus during social perception. J
Cogn Neurosci 16, 1706-1716.
Pelphrey, K.A., Morris, J.P., Michelich, C.R., Allison, T., & McCarthy, G. (2005). Functional anatomy of
biological motion perception in posterior temporal cortex: An fMRI study of eye, mouth and hand
movements. Cerebral Cortex, 15(12): 1866-76.
Pereira, F., & Mitchell, T. (2009). Machine learning classifiers and fMRI: A tutorial overview. NeuroImage.
Perner, J. (1991). Understanding the Representational Mind. Cambridge: MIT Press.
Perner, J. J., Aichhorn, M. M., Kronbichler, M. M., Staffen, W. W., & Ladurner, G. G. (2006). Thinking of
mental and other representations: the roles of left and right temporo-parietal junction. Social
Neuroscience, 1(3-4), 245-258. doi:10.1080/17470910600989896
Perner, J., & Lang, B. (1999). Development of theory of mind and executive control. Trends in cognitive
sciences, 3(9), 337-344.
Perner, J., & Leekam, S. (2008). The curious incident of the photo that was accused of being false: issues
of domain specificity in development, autism, and brain imaging. Quarterly Journal of Experimental
Psychology, 61(1), 76-89. doi:10.1080/17470210701508756
Perner, J., & Roessler, J. (2012). From infantsÒ to childrenÓs appreciation of belief. Trends in Cognitive
Sciences, 16(10), 519-525. doi:10.1016/j.tics.2012.08.004
170 of 179
Perner, J., Aichhorn, M., Kronbichler, M., Staffen, W., & Ladurner, G. (2006). Thinking of mental and other
representations: The roles of left and right temporo-parietal junction. Social Neuroscience 1(3-4):
245-58.
Perrett, D.I.D., Smith, P.A.P., Mistlin, A.J.A., Chitty, A.J.A., Head, A.S.A., Potter, D.D.D., Broennimann,
R.R., Milner, A.D.A., and Jeeves, M.A.M. (1985). Visual analysis of body movements by neurones in
the temporal cortex of the macaque monkey: a preliminary report. Behav Brain Res 16, 153-170.
Peter Miene, Bordiga, E., & Park, R. (1993). The Evaluation of Hearsay Evidence: A Social Psychological
Approach. In N. J. Castellan, Individual and Group Decision Making. Psychology Press.
Peter Miene, Park, R. C., & Bordiga, E. (1992). Juror Decision Making and the Evaluation of Hearsay
Evidence. 76 Minn . L. Rev, (683).
Peterson C., & Wellman, H. M. (2010). From fancy to reason: scaling deaf and hearing children's
understanding of theory of mind and pretence. British Journal of Developmental Psychology, 27(Pt 2),
297-310. doi:10.1348/026151008X299728
Peterson, & Siegal, M. (1999). Representing Inner Worlds: Theory of Mind in Autistic, Deaf, and Normal
Hearing Children. Psychological Science, 10(2), 126-129. doi:10.1111/1467-9280.00119
Peterson, C., Wellman, H. M., & Liu, D. (2005). Steps in theory-of-mind development for children with
deafness or autism. Child Development, 76(2), 502-517.
Peterson, J. L., & Webb, J. (2000). Factors influencing the development of a theory of mind in blind
c h i l d r e n . B r i t i s h J o u r n a l o f D e v e l o p m e n t a l P s y c h o l o g y, 1 8 ( 3 ) , 4 3 1 - 4 4 7 . d o i :
10.1348/026151000165788
Philip, R. C. M., Whalley, H. C., Stanfield, A. C., Sprengelmeyer, R., Santos, I. M., Young, A. W., et al.
(2010). Deficits in facial, body movement and vocal emotional processing in autism spectrum
disorders. Psychological Medicine, 40(11), 1919-1929. doi:10.1017/S0033291709992364
Pillow, B. H., Hill, V., Boyce, A., & Stein, C. (2000). Understanding inference as a source of knowledge:
children's ability to evaluate the certainty of deduction, perception, and guessing. Developmental
Psychology, 36(2), 169-179.
Platek, S. (2004). Where am I? The neurological correlates of self and other. Cognitive Brain Research,
19(2), 114-122. doi:10.1016/j.cogbrainres.2003.11.014
Platek, S., Keenan, J.P., Gallup Jr, G.G., & Mohamed, F.B. (2004). Where am I? The neurological
correlates of self and other. Cognitive Brain Research 19(2): 114-22.
Poldrack, R. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive
Sciences, 10(2), 59-63. doi:10.1016/j.tics.2005.12.004
Poldrack, R.A. (2006b). Region of interest analysis for fMRI. Social Cognitive and Affective Neuroscience
2(1): 67-70.
Poore, J. C., Pfeifer, J. H., Berkman, E. T., Inagaki, T. K., Welborn, B. L., & Lieberman, M. D. (2012).
Prediction-error in the context of real social relationships modulates reward system activity. Frontiers
in Human Neuroscience, 6, 1-11. doi:10.3389/fnhum.2012.00218
Puce, A.A., Allison, T.T., Bentin, S.S., Gore, J.C.J., and McCarthy, G.G. (1998). Temporal cortex activation
in humans viewing eye and mouth movements. J Neurosci 18, 2188-2199.
Pyers, J., & Senghas, A. (2009). Language Promotes False-Belief Understanding. Psychological Science,
20(7), 805.
Quian Quiroga, R., & Panzeri, S. (2009). Extracting information from neuronal populations: information
theory and decoding approaches. Nature Reviews Neuroscience, 10(3), 173-185. doi:10.1038/
nrn2578
171 of 179
R Core Team. (2013). R: a language and environment for statistical computing. Vienna: R Foundation for
Statistical Computing.
Rabin, J. S., Gilboa, A., Stuss, D. T., Mar, R. A., & Rosenbaum, R. S. (2010). Common and unique neural
correlates of autobiographical memory and theory of mind. Journal of Cognitive Neuroscience, 22(6),
1095-1111. doi:10.1162/jocn.2009.21344
Raizada, R. D. S., Tsao, F.-M., Liu, H.-M., Holloway, I. D., Ansari, D., & Kuhl, P. K. (2010). Linking brainwide multivoxel activation patterns to behaviour: Examples from language and math. NeuroImage,
51(1), 462-471. doi:10.1016/j.neuroimage.2010.01.080
Rao, R.P., and Ballard, D.H. (1999). Predictive coding in the visual cortex: a functional interpretation of
some extra-classical receptive-field effects. Nat Neurosci 2, 79-87.
Rayner, K., & Duffy, S. A. (1986). Lexical complexity and fixation times in reading: effects of word
frequency, verb complexity, and lexical ambiguity. Memory and Cognition, 14(3), 191-201.
Repacholi, B. M., & Gopnik, A. (1997). Early reasoning about desires: evidence from 14- and 18-montholds. Developmental Psychology, 33(1), 12-21.
Reynolds, J.H., and Heeger, D.J. (2009). The Normalization Model of Attention. Neuron 61, 168-185.
Robinson, E. J. (1994). What people say, what they think, and what is really the case: Children's
understanding of utterances as sources of knowledge. In C. Lewis & P. Mitchell, (pp. 355-381).
Children's early understanding of mind: Origins ƒ.
Robinson, E. J., & Whitcombe, E. L. (2003). Children's Suggestibility in Relation to their Understanding
about Sources of Knowledge. Child Development, 74(1), 48-62. doi:10.1111/1467-8624.t01-1-00520
Robinson, E. J., Champion, H., & Mitchell, P. (1999). Children's ability to infer utterance veracity from
speaker informedness. Developmental Psychology, 35(2), 535-546.
Robinson, E. J., Haigh, S. N., & Nurmsoo, E. (2008). Children's working understanding of knowledge
sources: Confidence in knowledge gained from testimony. Cognitive Development, 23(1), 105-118.
doi:10.1016/j.cogdev.2007.05.001
Rolls, E. T., & Treves, A. (2011). The neuronal encoding of information in the brain. Progress in
Neurobiology, 95(3), 448-490. doi:10.1016/j.pneurobio.2011.08.002
Rosel, J., Caballer, A., Jara, P., & Oliver, J. C. (2005). Verbalism in the Narrative Language of Children
Who Are Blind and Sighted. Journal of Visual Impairment and Blindness, 99, 413-425.
Ross, D. R., Ceci, S. J., Dunning, D., & Toglia, M. P. (1994). Unconscious transference and mistake
identity: When a witness misidentifies a familiar with innocent person. Journal of Applied Psychology,
79(6), 918-930.
Ross, L., Greene, D., & House, P. (1977). The false consensus effect: An egocentric bias in social
perception and attribution processes. J Exp Soc Psychol, 13, 279 -301.
Rushworth, M. F., Mars, R. B., & Sallet, J. (2013). Are there specialized circuits for social cognition and
are they unique to humans?. Curr Opin Neurobiol 23(3) 436-442
Samson, D., Apperly, I. A., Braithwaite, J. J., Andrews, B. J., & Bodley Scott, S. E. (2010). Seeing it their
way: Evidence for rapid and involuntary computation of what other people see. Journal of
Experimental Psychology: Human Perception and Performance, 36(5), 1255-1266. doi:10.1037/
a0018729
Samson, D., Apperly, I.A., & Humphreys, G.W. (2007). Error analyses reveal contrasting deficits in 'theory
of mind': Neuropsychological evidence from a 3-option false belief task. Neuropsychologia 45(11):
2561-9.
Santos, L., Flombaum, J., and Phillips, W. (2013). The Evolution of Human Mindreading. Evol Cogn
Neurosci 362, 659-669.
172 of 179
Sapountzis, P., Schluppeck, D., Bowtell, R., & Peirce, J. W. (2010). A comparison of fMRI adaptation and
multivariate pattern classification analysis in visual cortex. Neuroimage, 49(2), 1632-1640.
Saxe, R.R. (2005). Against simulation: the argument from error. Trends in Cognitive Sciences, 9(4),
174-179. doi:10.1016/j.tics.2005.01.012
Saxe, R.R. (2006). Uniquely human social cognition. Curr Opin Neurobiol 16, 235-239.
Saxe, R.R. (2013). The new puzzle of theory of mind development. In M. Banaji & S. A. Gelman,
Navigating the Social World What Infants, Children, and Other Species Can Teach Us. Navigating the
Social World: What Infants.
Saxe, R.R., & Kanwisher, N. (2003). People thinking about thinking people:: The role of the temporoparietal junction in. NeuroImage, 19(4), 1835-1842.
Saxe, R.R., & Offen, S. (2010). Seeing ourselves: What vision can teach us about metacognition. In: G.
Dimaggio, & P.H. Lysaker (Eds), Metacognition and Severe Adult Mental Disorders: From basic
research to treatment (pp. 13-29). New York: Taylor & Francis.
Saxe, R.R., & Powell, L. J. (2006). It's the thought that counts: specific brain regions for one component
of theory of mind. Psychological Science, 17(8), 692-699. doi:10.1111/j.1467-9280.2006.01768.x
Saxe, R.R., & Wexler, A. (2005). Making sense of another mind: The role of the right temporo-parietal
junction. Neuropsychologia, 43(10), 1391-1399. doi:10.1016/j.neuropsychologia.2005.02.013
Saxe, R.R., Moran, J.M., Scholz, J., & Gabrieli, J. (2006a). Overlapping and non-overlapping brain
regions for theory of mind and self reflection in individual subjects. Social Cognitive and Affective
Neuroscience 1(3): 229-34.
Saxe, R.R., Schulz, L. E., & Jiang, Y. V. (2006). Reading minds versus following rules: Dissociating theory
of mind and executive control in the brain. Social Neuroscience, 1(3-4), 284-298. doi:
10.1080/17470910601000446
Saxe, R.R., Whitfield-Gabrieli, S., Scholz, J., & Pelphrey, K.A. (2009). Brain regions for perceiving and
reasoning about other people in school-aged children. Child Development, 80(4): 1197-209.
Saxe, R.R., Xiao, D.-K., Kovacs, G., Perrett, D.I., and Kanwisher, N. (2004). A region of right posterior
superior temporal sulcus responds to observed intentional actions. Neuropsychologia 42, 12-12.
Saygin, A.P., Chaminade, T., Ishiguro, H., Driver, J., and Frith, C. (2012). The thing that should not be:
predictive coding and the uncanny valley in perceiving human and humanoid robot actions. Soc Cogn
Affect Neurosci 7, 413-422.
Schick, B., de Villiers, P., de Villiers, J., & Hoffmeister, R. (2007). Language and theory of mind: A study of
deaf children. Child Development, 78(2), 376-396.
Schiller, D., Freeman, J.B., Mitchell, J.P., Uleman, J.S., and Phelps, E.A. (2009). A neural mechanism of
first impressions. Nat Neurosci 12, 508-514.
Schneider, W., Schumann-Hengsteler, R., & Sodian, B. (Eds.). (2014). Young children's cognitive
development: Interrelationships among executive functioning, working memory, verbal ability, and
theory of mind. Psychology Press.
Schnell, K., Bluschke, S., Konradt, B., & Walter, H. (2010). Functional relations of empathy and
mentalizing: An fMRI study on the neural basis of cognitive empathy. NeuroImage, 1-41. doi:10.1016/
j.neuroimage.2010.08.024
Scholz, J., Triantafyllou, C., Whitfield-Gabrieli, S., Brown, E. N., & Saxe, R.R. (2009). Distinct Regions of
Right Temporo-Parietal Junction Are Selective for Theory of Mind and Exogenous Attention. PLoS
ONE, 4(3), e4869. doi:10.1371/journal.pone.0004869
Schultz, W. (2010). Dopamine signals for reward value and risk: basic and recent data. Behav Brain
Funct 6, 24.
173 of 179
Schultz, W., Dayan, P., and Montague, P.R. (1997). A neural substrate of prediction and reward. Science
275, 1593-1599.
Schurz, M., Radua, J., Aichhorn, M. M., Richlan, F., & Perner, J. (2014). Fractionating theory of mind: a
meta-analysis of functional brain imaging studies. Neuroscience & Biobehavioral Reviews, 42, 9-34.
doi:10.1016/j.neubiorev.2014.01.009
Scofield, J., & Behrend, D. A. (2008). Learning words from reliable and unreliable speakers. Cognitive
Development, 23(2), 278-290. doi:10.1016/j.cogdev.2008.01.003
Senior, C.C., Barnes, J.J., Giampietro, V.V., Simmons, A.A., Bullmore, E.T.E., Brammer, M.M., and David,
A.S.A. (2000). The functional neuroanatomy of implicit-motion perception or 'representational
momentum.' Curr Biol 10, 7-7.
Senju, A., Johnson, M.H., and Csibra, G. (2006). The development and neural basis of referential gaze
perception. Soc Neurosci 1, 220-234.
Seo, H.H., and Lee, D.D. (2012). Neural basis of learning and preference during social decision-making.
Curr Opin Neurobiol 22, 990-995.
Serences, J. T., & Boynton, G. M. (2007). Feature-Based Attentional Modulations in the Absence of Direct
Visual Stimulation. Neuron, 55(2), 301-312. doi:10.1016/j.neuron.2007.06.015
Serences, J.T., Shomstein, S., Leber, A.B., Golay, X., Egeth, H.E., & Yantis, S. (2005). Coordination of
voluntary and stimulus-driven attentional control in human cortex. Psychological Science 16:114-22.
Serre, T., Wolf, L., & Poggio, T. (2005, June). Object recognition with features inspired by visual cortex. In
Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference on
(Vol. 2, pp. 994-1000). IEEE.
Seung, H. S., & Sompolinsky, H. (1993). Simple models for reading neuronal population codes, 90(22),
10749-10753. doi:10.1073/pnas.90.22.10749
Shamay-Tsoory, S., Tomer, R., Berger, B.D., Goldsher, D., & Aharon-Peretz, J. (2005). Impaired affective
theory of mind is associated with right ventromedial prefrontal damage. Cognitive and Behavioral
Neurology 18(1): 55-67.
Shamay-Tsoory, S.G., & Aharon-Peretz, J. (2007). Dissociable prefrontal networks for cognitive and
affective theory of mind: a lesion study. Neuropsychologia 45(13): 3054-67.
Shamir, M., & Sompolinsky, H. (2006). Implications of neuronal diversity on population coding. Neural
Computation, 18(8), 1951-1986. doi:10.1162/neco.2006.18.8.1951
Shepard, R. N. (1980). Multidimensional scaling, tree-fitting, and clustering.Science, 210(4468), 390-398.
Shepard, R. N., & Cooper, L. A. (1992). Representation of Colors in the Blind, Color-Blind, and Normally
Sighted. Psychological Science, 3(2), 97-104. doi:10.1111/j.1467-9280.1992.tb00006.x
Shibata, H., Inui, T., & Ogawa, K. (2011). Understanding interpersonal action coordination: an fMRI study.
Exp Brain Res, 211(3-4), 569-579.
Shultz, S., Lee, S.M., Pelphrey, K., and McCarthy, G. (2011). The posterior superior temporal sulcus is
sensitive to the outcome of human and non-human goal-directed actions. Soc Cogn Affect Neurosci
6, 602-611.
Simanova, I., Hagoort, P., Oostenveld, R., & van Gerven, M. A. (2014). Modality-independent decoding of
semantic information from the human brain.Cerebral Cortex, 24(2), 426-434.
Singer, T., Ben Seymour, O'Doherty, J., Kaube, H., Dolan, R. J., & Frith, C. D. (2004). Empathy for pain
involves the affective but not sensory components of pain. Science, 303(5661), 1157-1162. doi:
10.1126/science.1093535
174 of 179
Slessor, G., Phillips, L.H., & Bull, R. (2007). Exploring the specificity of age-related differences in theory of
mind tasks. Psychology and Aging, 22(3): 639-43.
Smirnova, A. (2012). Evidentiality in Bulgarian: Temporality, Epistemic Modality, and Information Source.
Journal of Semantics. doi:10.1093/jos/ffs017
Smith, A. T., Kosillo, P., & Williams, A. L. (2011). The confounding effect of response amplitude on MVPA
performance measures. NeuroImage, 56(2), 525-530. doi:10.1016/j.neuroimage.2010.05.079
Smith, V. L., & Ellsworth, P. C. (1987). The social psychology of eyewitness accuracy: Misleading
questions and communicator expertise. Journal of Applied Psychology, 72(2), 294-300. doi:
10.1037/0021-9010.72.2.294
Sommer, M., Döhnel, K., Sodian, B., Meinhardt, J., Thoermer, C., & Hajak, G. (2007). Neural correlates of
true and false belief reasoning. NeuroImage, 35(3), 1378-1384. doi:10.1016/j.neuroimage.
2007.01.042
Sommer, M., Rothmayr, C., Döhnel, K., Meinhardt, J., Schwerdtner, J., Sodian, B., & Hajak, G. (2010).
How should I decide? The neural correlates of everyday moral reasoning. Neuropsychologia, 48(7):
2018-26.
Southgate, V., Senju, A., & Csibra, G. (2007). Action anticipation through attribution of false belief by 2year-olds. Psychological Science 18(7): 587-92.
Spiers, H. J., & Maguire, E. A. (2007). Decoding human brain activity during real-world experiences.
Trends in Cognitive Sciences, 11(8), 356-365. doi:10.1016/j.tics.2007.06.002
Spiers, H.J., & Maguire, E.A. (2006). Thoughts, behaviour, and brain dynamics during navigation in the
real world. NeuroImage, 31(4): 1826-40.
Spratling, M.W. (2008). Reconciling Predictive Coding and Biased Competition Models of Cortical
Function. Front Comp Neurosci 2(4), 1-8.
Spratling, M.W. (2010). Predictive Coding as a Model of Response Properties in Cortical Area V1. J
Neurosci 30, 3531-3543.
Spunt, R. P., & Lieberman, M. D. (2012). An integrative model of the neural systems supporting the
comprehension of observed emotional behavior. NeuroImage, 59(3), 3050-3059. doi:10.1016/
j.neuroimage.2011.10.005
Spunt, R., Satpute, A.B., & Lieberman, M. (2011). Identifying the what, why, and how of an observed
action: an fMRI study of mentalizing and mechanizing during action observation. Journal of Cognitive
Neuroscience 23(1): 63-74.
Spunt, R.P., & Lieberman, M.D. (2012). An integrative model of the neural systems supporting the
comprehension of observed emotional behavior. NeuroImage 59(3), 3050.
Steinwart, I., & Christmann, A. (2008). Support Vector Machines. New York, NY: Springer Science &
Business Media. doi:10.1007/978-0-387-77242-4
Stelzer, J., Chen, Y., & Turner, R. (2013). Statistical inference and multiple testing correction in
classification-based multi-voxel pattern analysis (MVPA): Random permutations and cluster size
control. NeuroImage, 65, 69-82. doi:10.1016/j.neuroimage.2012.09.063
Stich, S., & Nichols, S. (1992). Folk Psychology: Simulation or Tacit Theory? Mind & Language, 7(1-2),
35-71.
Summerfield, C., Trittschuh, E.H., Monti, J.M., Mesulam, M.M., and Egner, T. (2008). Neural repetition
suppression reflects fulfilled perceptual expectations. Nat Neurosci 11, 1004-1006.
Tamir, D.I. & Mitchell, J.P. (2010). Neural correlates of anchoring-and-adjustment during mentalizing.
Proceedings of the National Academy of Sciences, 107(24): 10827.
175 of 179
Tenenbaum, J.B., Kemp, C., Griffiths, T.L., and Goodman, N.D. (2011). How to Grow a Mind: Statistics,
Structure, and Abstraction. Science 331, 1279-1285.
Tesink, C. M. J. Y., Buitelaar, J. K., Petersson, K. M., van der Gaag, R. J., Kan, C. C., Tendolkar, I., &
Hagoort, P. (2009). Neural correlates of pragmatic language comprehension in autism spectrum
disorders. Brain, 132(7), 1941-1952. doi:10.1093/brain/awp103
Tobler, P.N., Fiorillo, C.D., and Schultz, W. (2005). Adaptive coding of reward value by dopamine neurons.
Science 307, 1642-1645.
Todorovic, A., & de Lange, F. P. (2012). Repetition suppression and expectation suppression are
dissociable in time in early auditory evoked fields. J Neurosci, 32(39), 13389-13395.
Todorovic, A., van Ede, F., Maris, E., and de Lange, F.P. (2011). Prior Expectation Mediates Neural
Adaptation to Repeated Sounds in the Auditory Cortex: An MEG Study. J Neurosci 31, 9118-9123.
Toglia, M. P., Ross, D. F., Ceci, S. J., & Hembrooke, H. (1992). The Suggestibility of Children's Memory:
A Social-Psychological and Cognitive Interpretation. In Development of Long-Term Retention (pp.
217-241). New York, NY: Springer New York. doi:10.1007/978-1-4612-2868-4-7
Treue, S., and MartÕnez-Trujillo, J.C. (1999). Feature-based attention influences motion processing gain
in macaque visual cortex. Nature 399, 575-579.
Tsakiris, M., Costantini, M., & Haggard, P. (2008). The role of the right temporo-parietal junction in
maintaining a coherent sense of one's body. Neuropsychologia 46(12): 3014-18.
Tudusciuc, O., & Nieder, A. (2007). Neuronal population coding of continuous and discrete quantity in the
primate posterior parietal cortex. Proceedings of the National Academy of Sciences of the United
States of America, 104(36), 14513-14518. doi:10.1073/pnas.0705495104
Turk-Browne, N.B., Yi, D.-J., and Chun, M.M. (2006). Linking implicit and explicit memory: common
encoding factors and shared representations. Neuron 49, 917-927.
Tusche, A., Bode, S., & Haynes, J. D. (2010). Neural Responses to Unattended Products Predict Later
Consumer Choices. Journal of Neuroscience, 30(23), 8024-8031. doi:10.1523/JNEUROSCI.
0064-10.2010
Urwin, C. (1983). Dialogue and cognitive functioning in the early language development of three blind
children. In A. E. Mills (Ed.), Language acquisition in the blind child: Normal and deficient (pp.
142-161). San Diego, CA: College-Hill Press.
Uttal, W.R. (2011). Mind and Brain: A Critical Appraisal of Cognitive Neuroscience. Cambridge: MIT Press.
Van Overwalle, F. (2008). Social cognition and the brain: A meta-analysis. Human Brain Mapping 30(3):
829-58.
Van Overwalle, F., & Van Overwalle, F. (2009). Social cognition and the brain: a meta-analysis. Human
Brain Mapping, 30(3), 829-858. doi:10.1002/hbm.20547
Vander Wyk, B.C., Hudac, C.M., Carter, E.J., Sobel, D.M., and Pelphrey, K.A. (2009). Action
understanding in the superior temporal sulcus region. Psychol Sci 20, 771-777.
Vapnik, V. N., & Vapnik, V. (1998). Statistical learning theory, 2.
Visser, M., Jefferies, E., Embleton, K. V., & Ralph, M. A. L. (2012). Both the middle temporal gyrus and
the ventral anterior temporal area are crucial for multimodal semantic processing: distortion-corrected
fMRI evidence for a double gradient of information convergence in the temporal lobes. Journal of
Cognitive Neuroscience, 24(8), 1766-1778.
Vogeley, K., Bussfeld, P., Newen, A., Herrmann, S., Happe, F., Falkai, P., ... & Zilles, K. (2001). Mind
reading: neural mechanisms of theory of mind and self-perspective. Neuroimage, 14(1), 170-181.
176 of 179
Vogeley, K., May, M., Ritzl, A., Falkai, P., Zilles, K., & Fink, G.R. (2004). Neural correlates of first-person
perspective as one constituent of human self-consciousness. Journal of Cognitive Neuroscience 16,
817-27.
Vuilleumier, P., Armony, J.L., Driver, J., & Dolan, R.J. (2001). Effects of attention and emotion on face
processing in the human brain:: an event-related fMRI study. Neuron, 30(3): 829-41.
Wacongne, C., Changeux, J.P., and Dehaene, S. (2012). A Neuronal Model of Predictive Coding
Accounting for the Mismatch Negativity. J Neurosci 32, 3665-3678.
Wagner, D.D., Kelley, W.M., & Heatherton, T.F. (2011). Individual differences in the spontaneous
recruitment of brain regions supporting mental state understanding when viewing natural social
scenes. Cerebral Cortex, 21(12): 2788-96.
Wallis, J. D., Anderson, K. C., & Miller, E. K. (2001). Single neurons in prefrontal cortex encode abstract
rules. Nature, 411(6840), 953-956.
Walter, H., Schnell, K., Erk, S., Arnold, C., Kirsch, P., Esslinger, C., et al. (2010). Effects of a genome-wide
supported psychosis risk variant on neural activation during a theory-of-mind task. Molecular
Psychiatry, 16(4), 462-470. doi:doi:10.1038/mp.2010.18
Walther, D. B., Beck, D. M., & Fei-Fei, L. (2012). To err is human: Correlating fMRI decoding and
behavioral errors to probe the neural representation of natural scene categories. In N. Kriegeskorte &
G. Kreiman, (pp. 391-415). Visual population codes-Toward a common multivariate framework for cell
recording and functional imaging.
Wang, A. T. (2006). Neural basis of irony comprehension in children with autism: the role of prosody and
context. Brain, 129(4), 932-943. doi:10.1093/brain/awl032
Weiner, K.S., Sayres, R., Vinberg, J., and Grill-Spector, K. (2010). fMRI-Adaptation and Category
Selectivity in Human Ventral Temporal Cortex: Regional Differences Across Time Scales. J
Neurophysiol 103, 3349-3365.
Welch-Ross, M. K. (1999). Interviewer Knowledge and Preschoolers' Reasoning About Knowledge States
Moderate Suggestibility. Cognitive Development.
Welch-Ross, M. K., Diecidue, K., & Miller, S. A. (1997). Young children's understanding of conflicting
mental representation predicts suggestibility. Developmental Psychology, 33(1), 43-53.
Wellman, H. M., Cross, D., & Watson, J. (2001). Meta-analysis of theory-of-mind development: The truth
about false belief. Child Development, 655-684.
Wellman, H., Cross, D., & Watson, J. (2001). Meta-analysis of theory-of-mind development: The truth
about false belief. Child Development, 72(3). 655-84.
Whitcombe, E. L., & Robinson, E. J. (2000). Children's decisions about what to believe and their ability to
report the source of their belief. Cognitive Development, 15(3), 329-346. doi:10.1016/
S0885-2014(00)00033-2
Whitfield-Gabrieli, S., Moran, J.M., Nieto-Castañón, A., Triantafyllou, C., Saxe, R.R., & Gabrieli, J.D.
(2011). Associations and dissociations between default and self-reference networks in the human
brain. NeuroImage, 55(1): 225-32.
Wimmer, H., & Perner, J. (1983). Beliefs about beliefs: Representation and constraining function of wrong
beliefs in young children's understanding of deception. Cognition, 13(1), 103-128.
Winston, J.S., Henson, R.N., Fine-Goulden, M.R., and Dolan, R.J. (2004). fMRI-adaptation reveals
dissociable neural representations of identity and expression in face perception. J Neurophysiol 92,
1830-1839.
Woolfe, T., & Want, S. (2002). Signposts to development: Theory of mind in deaf children. Child
Development, 73(3), 768-778.
177 of 179
Woolfolk, R. L., Doris, J. M., & Darley, J. M. (2006). Identification, situational constraint, and social
cognition: studies in the attribution of moral responsibility. Cognition, 100(2), 283-301. doi:10.1016/
j.cognition.2005.05.002
Worsley, K.J., Evans, A.C., Marrett, S., & Neelin, P. (1992). A three- dimensional statistical analysis for
CBF activation studies in human brain. Journal of Cerebral Blood Flow & Metabolism 12(6): 900-18.
Yamada, M., Camerer, C. F., Fujie, S., Kato, M., Matsuda, T., Takano, H., et al. (2012). Neural circuits in
the brain that are activated when mitigating criminal sentences. Nature Communications, 3, 1-6. doi:
10.1038/ncomms1757
Yamada, Makiko, Colin F. Camerer, Saori Fujie, Motoichiro Kato, Tetsuya Matsuda, Harumasa Takano,
Hiroshi Ito, Tetsuya Suhara, and Hidehiko Takahashi. (2013) Neural circuits in the brain that are
activated when mitigating criminal sentences. Nature 3, 759.
Yang, D., Baillargeon, R., & Baillargeon, R. (2013). Difficulty in Understanding Social Acting (but not False
Beliefs) Mediates the Link Between Autistic Traits and Ingroup Relationships. Journal of Autism and
Developmental Disorders, 1-24.
Yoshida, W., Seymour, B., Friston, K.J., & Dolan, R.J. (2010). Neural mechanisms of belief inference
during cooperative games. Journal of Neuroscience 30(32): 10744-51.
Young, L.L., & Phillips, J. (2011). The paradox of moral focus. Cognition, 119(2), 166-178. doi:10.1016/
j.cognition.2011.01.004
Young, L.L., & Saxe, R.R. (2008). The neural basis of belief encoding and integration in moral judgment.
NeuroImage, 40(4), 1912-1920. doi:10.1016/j.neuroimage.2008.01.057
Young, L.L., & Saxe, R.R. (2009a). An FMRI investigation of spontaneous mental state inference for
moral judgment. Journal of Cognitive Neuroscience, 21(7), 1396-1405. doi:10.1086/591811
Young, L.L., & Saxe, R.R. (2009b). Innocent intentions: A correlation between forgiveness for accidental
harm and neural activity. Neuropsychologia, 47(10), 2065-2072. doi:10.1016/j.neuropsychologia.
2009.03.020
Young, L.L., & Waytz, A. (2013). Mind attribution is for morality. In S. L. Baron-Cohen, H. Tager-Flusberg,
& M . V. L o m b a r d o , U n d e r s t a n d i n g O t h e r M i n d s . O x f o r d U n i v e r s i t y P r e s s . d o i :
10.1177/0956797612472343
Young, L.L., Camprodon, J. A., Hauser, M., Pascual-Leone, A., & Saxe, R.R. (2010a). Disruption of the
right temporoparietal junction with transcranial magnetic stimulation reduces the role of beliefs in
moral judgments. Proceedings of the National Academy of Sciences of the United States of America,
107(15), 6753-6758. doi:10.1073/pnas.0914826107
Young, L.L., Cushman, F., Hauser, M., & Saxe, R.R. (2007). The neural basis of the interaction between
theory of mind and moral judgment. Proceedings of the National Academy of Sciences 104(20):
8235-40.
Young, L.L., Dodell-Feder, D., & Saxe, R.R. (2010b). What gets the attention of the temporo-parietal
junction? An fMRI investigation of attention and theory of mind. Neuropsychologia, 48(9), 2658-2664.
doi:10.1016/j.neuropsychologia.2010.05.012
Young, L.L., Nichols, S., & Saxe, R.R. (2010c). Investigating the Neural and Cognitive Basis of Moral
Luck: It's Not What You Do but What You Know. Review of Philosophy and Psychology, 1(3),
333-349. doi:10.1007/s13164-010-0027-y
Zahn, R., Moll, J., Krueger, F., Huey, E. D., Garrido, G., & Grafman, J. (2007). Social concepts are
represented in the superior anterior temporal cortex.Proceedings of the National Academy of
Sciences, 104(15), 6430-6435.
Zaitchik, D. (1990). When representations conflict with reality: The preschooler's problem with false
beliefs and. Cognition 35: 41-68.
178 of 179
Zaitchik, D., Koff, E., Brownell, H., Winner, E., & Albert, M. (2006). Inference of beliefs and emotions in
patients with Alzheimer. Neuropsychology 20, 11-20.
Zaitchik, D., Walker, C., Miller, S., LaViolette, P., Feczko, E., & Dickerson, B. C. (2010). Mental state
attribution and the temporoparietal junction: An fMRI study comparing belief, emotion, and perception.
Neuropsychologia, 48(9), 2528-2536. doi:10.1016/j.neuropsychologia.2010.04.031
Zaki, J., and Mitchell, J. P. (2011). Equitable decision making is associated with neural markers of intrinsic
value. Proc Natl Acad Sci U S A, 108(49), 19761-19766.
Zaragoza, M. S., & Lane, S. M. (1994). Source misattributions and the suggestibility of eyewitness
memory. Journal of Experimental Psychology-Learning Memory and Cognition, 20(4), 934-945.
Zhu, L., Mathewson, K.E., and Hsu, M. (2012). Dissociable neural representations of reinforcement and
belief prediction errors underlie strategic learning. Proc Natl Acad Sci U S A 109, 1419-1424.
179 of 179