Epistemic norms in procedural and analytic

Theory of Mind, Simulation and Meta-Cognition
Laureate's Colloquium with Josef Perner
Epistemic norms in procedural
and analytic metacognition
Joëlle Proust
Institut Jean-Nicod
1. On Perner’s 3 criteria for « minimeta »
2. 2 pernicious problems inherent to potential evidence
for non-human « metacognition » (Perner, 2012)
3. Addressing problem 1: knowing-that versus expressing
4. Addressing problem 2: subjective uncertainty vs world
5. Metacognitive norms in procedural and analytic
On Perner’s 3 criteria for
« minimeta »
Josef Perner’s (2012) conception of
• In analogy to metarepresentation as understood in Perner
(1991), metacognition is defined as cognition about
cognitions as cognitions, which implies some understanding
of what makes cognitions what they are, namely their
representational content. (FM, 96)
Cognitive levels (Perner, 2012)
Scope of metacognition ?
Genuine metacognition: cognition about
cognition as cognition
? Miniméta (or special cognition):
• Cognition about cognition without recursion:
intentional content not involved.
3 criteria for MiniMeta
Necessity: is the component cognition that makes behaviour
intuitively ‘meta-’ necessary for the behaviour to occur?
Directionality: “MiniMetacognition is cognition that goes beyond
ordinary object-level cognition in the direction of recursive
metacognition”. Implicit admission of ignorance OK.
Exclusivity: MiniMetacognition should only be needed for
behaviour that is intuitively metacognitive.
Necessity: a shared
methodological constraint
• a process needs to help an organism
ameliorate its first-order performance to
qualify as metacognitive
Robert Hampton (2009): Objective markers for
metacognitive behavior:
1. There must be a primary behavior that can
be scored for its accuracy.
2. Variation in performance (i.e. uncertainty
about outcome) must be present.
3. A secondary behavior, whose goal is to
regulate the primary behavior, must be
elicited in the animal.
4. This secondary behavior must be shown to
benefit performance in the primary task (for
example, animals must decline tests that
they would otherwise have failed).
• The notion of “directionality” is a metaphor for
what goes beyond first-level cognition.
• The term may be misleading, as it suggests that
metacognition should more or less anticipate
second-level cognition, where anticipation of
ignorance, understood as an ability to
metarepresent that one does not know.
• « Directionality » might be re-stated in neutral
– using information that is not directly presented in
the perceptual content of a first-order task
– In order to monitor and control a cognitive action
(i.e., aiming to acquire an epistemic property)
• MiniMetacognition should only be needed for
behaviour that is intuitively metacognitive
This condition is a requirement for any functionalist
analysis of the mental.
If metacognition is a natural kind, then it should have its
inputs, triggering conditions, needs,
– informational use,
– Causal mechanisms
– Downstream decisions.
• Proust (2007) criticized for failing the exclusivity
• Rejoinder: Proust’s (2007) claim is that:
– the types of control of physical and mental agency have
analogous structural descriptions (comparator, feedback,
– but they deeply differ in the information being used, in
their phylogenetic distribution, their specific mechanisms
• « Let us start with a non metacognitive case such as
evaluating a possible bodily action. »
(Proust, 2007, Section III, 1)
• For a discussion of the differences between metacognition
and meta-action, see also: Proust (2008), (2009), (in print)
• For the differences in downstream decisions, see Proust
(2012a, 2012b).
Symmary of the discussion about
the 3 criteria for MM
• Wide agreement on conditions 1 and 3, i.e. necessity
and exclusivity.
• « Directionality » is an acceptable constraint once
re-stated in neutral terms.
« A pair of pernicious problems plagues
potential evidence for non-human
« metacognition » (Perner, 2012)
« A pair of pernicious problems »
Difficulty of experimentally prying apart:
1.Being in a state S (of knowing, believing),
etc. from Knowing that one is in a state S . E.g:
being uncertain vs knowing that one is uncertain.
2.Representing a state of the world, from
representing an inner state caused by a state of the
world. E.g.: finding a task difficult/judging oneself
as finding a task difficult
Example of problem 1: Call & Carpenter
• In the knowledge condition, chimpanzees, who know
where the bait is, go for it and do not look around.
• In the partial ignorance condition, chimps are looking into
the tubes.
• It is no evidence of metacognition: these behaviours can
be governed by the degree of the chimpanzee’s
knowledge/ignorance without any recursive cognitions
about his degree of knowledge.
Example of problem 2: Opt out paradigm
Smith et al. (1997, 2006)
The decision to skip can be based
• Either: on realizing that this is a difficult problem
(a cognition about an external state of the world)
• Or: on realizing that I am uncertain (a clear
Perner, (2012)
Addressing Problem 1 : why
should the control of one’s
cognition be based on a knowthat to count as metacognitive?
Being uncertain vs knowing that one is
First argument against the « know that »
•In many cognitive domains, « feeling » is a
recognized independent source of information,
that can reliably influence rational decision.
•Having a feeling about conducting some activity
does not require knowing that one has it.
•Feelings can directly orient subjects’ decisions to
Being uncertain vs knowing that one is
• « Knowing that » involves a propositional
• Feeling certain/uncertain does not need to
involve a propositional content to be « about » a
state of mind (more on this later)
• Feelings can be ingredients in know-hows and
• They causally influence decisions based on one’s
experienced degree of certainty/uncertainty
Accumulation of evidence as the information on
which noetic feelings depend
• Between behaviour and propositional thought,
an additional type of information is based on
dynamic features in the neural vehicle:
In perceptual or memorial uncertainty:
–Onset of neural activity
–Neural coherence in the responses
–Dynamics of convergence to a decision.
Knowing-That / Expressing
• Second argument against the « know that »
Work in formal semantics shows that a
metarepresentational view of probabilities
applied to our credal states does not work.
Yalcin (2011), Leitgeb (2012),
Seth Yalcin (2007, 2011): credal
• Epistemic modals
« When expressing one’s state of belief by saying ‘It
might be raining’, one is not expressing a
proposition one believes. Rather, one is expressing
the openness of one’s state of belief – a property of
the state of belief ».
• The language of probability
« Probability sentences semantically express conditions
on states of information which are not reducible to
conditions on possible worlds (ways the world might
be, or truth-conditions in the usual sense).
Indicative conditionals (e.g. « If Oswald did
not kill Kennedy, someone else did »)
• Suppositionalists about indicative conditionals are
claiming that “indicative conditionals express an
agent’s state of mind without saying that the agent’s
state of mind is so-and-so.”
• Accordingly, one might expect an explication of
‘metacognitive’’ to at least leave open that
metacognitive states and processes are about an
agent’s internal states or processes in the sense of
expressing these states or processes, in the same
sense in which ‘Yippie!’ expresses a positive emotional
state, without stating that the state is so-and-so
Leitgeb (2012), 264
Expressivism: a new sense of ‘aboutness’
• While states of mind may be reported, or
attributively accessed through doxastic (e.g.
introspective or interpretive) means, they can
also be made accessible in an immediate, non
doxastic way. (Gibbard, 1990, Bar-On 2004).
• Expressing one's thought in this direct way
can be done even by subjects unable to
represent the fact that they have mental
states, such as non-humans and human
young children.
Being in a state/Knowing one is in a
state/ Expressing a state
• A third term: expressing one’s being in a state
by having the corresponding noetic feeling.
• This feeling allows monitoring one’s state and
controlling it without having a conceptual
representation of that state.
Addressing Problem 2
• “Being uncertain is, like being hot, an inner
state. It is, unlike being hot, also a cognitive
state—but it is not a metacognitive state
(..)So why would learning to skip difficult test
items be evidence for metacognition?”
(Perner, 2012, 100)
Representing a state of the world or an
inner state caused by a state of the world?
• Uncertainty is the property of a belief, that agents
express through noetic feelings.
• A major functional difference between representing
that P, and being uncertain whether P:
– Representing that P (in its simpler form) is
sensitive to world variance, but insensitive to the
agent ’s own success in the task
– being U that P is sensitive to the latter, even
when no trial-by-trial reward is currently
The case of « difficult »
• « Difficult » in the epistemic sense is only applied
reliably when the subject has been trained in a
cognitive task.
• School children start being overconfident in their
memory/perception, and progressively learn how to
distinguish « easy guess » from « accurate answer ».
So do monkeys.
• Ease of processing can be seen as an epistemic
standard or norm, that agents use early on (e.g.
feeling of familiarity).
Oppenheimer (2008), Brinck & Lilienfors (2012)
• On this view, it is the evolutionary and
developmental basis for all the forms of
Metacognition as a procedural and
as an analytic capacity and its
What are epistemic norms?
• The normative feature of epistemic norms derives
from the structure of action being polarized (success
vs failure)
• A given norm is what regulates self-evaluation , i.e.
action monitoring, in a task-specific way.
• Norms can be epistemic, moral, rational, social,
• Epistemic norms are those that regulate selfevaluation in cognitive actions.
Why should metacognition involve
• Granting that metacognition has as its
function to allow an agent to make correct
decisions, metacognitive decisions implicitly
express sensitivity to various epistemic
norms, such as validity in perception,
accuracy or exhaustiveness in memory,
coherence and informativeness in
Epistemic norms
• Fluency or Intelligibility (perceptual
judgment, epistemic vigilance)
• Accuracy (memory, reasoning)
• Comprehensiveness or exhaustiveness
(memory, reasoning)
• Coherence (fiction, demonstrative reasoning)
• Consensus (negotiation)
• Relevance (conversation)
• Plausibility …/
An evolutionary hypothesis
Metacognition has evolved
• from a procedural type of cognitive control
where sensitivity to fluency is driving
perceptual and memorial decisions
• To an analytic type of cognitive control,
where sensitivity to fluency can be overruled
by externalized standards.
The paradox of fluency as an epistemic
• Josef Perner’s doubts about procedural
metacognition reflect the paradox of what
constitutes the basis of our critical minds.
• Sensitivity to fluency is based on an activity –
dependent information: the dynamic features of
the neural vehicle underlying a given task, rather
than on the basis of the associated cognitive
Nonconceptual representations of FBS carry
information through dynamic properties
• Comparative fluency is the property, for a
stimulus, to be processed more or less quickly
and adequately, with respect to what is
expected, in a kind of task.
• This property is a gradient on a normative
scale: it works as an indicator for what
successful processing should be like, for a task
in a context.
Objection: why should fluency count as an
epistemic norm?
Fluency qualifies as an epistemic norm
•Because agents sensitive to it can reliably assess
their perceptual and memorial decisions for
correctness (“knowledge-conducive”)
•Because sensitivity to it can be adaptively trained
to responding to new thresholds/ conditions.
•Cues for Fluency, however, are often misapplied in
cases where ease of processing cannot be profitably
used to predict correct answer.
Fluency in children’s early metacognitiion
• Klow and Rohwer (2012), and Beck et al. (2012)
found that children have more difficulty when they
can represent or imagine an answer (partial
ignorance or epistemic uncertainty conditions)
rather than when no image comes to their minds
(full ignorance or physical uncertainty conditions).
Fluency central in children’s early
2 interpretations:
1.Being able to imagine a possible outcome might
impair children’s ability to make [analytic]
metacognitive evaluations.
2.Being able to image a possible outcome might
allow children to make an [experience-based]
metacognitive evaluation, based on the fluency of
their quickly coming up with a possible answer.
Fluency in children’s early metacognitiion
“That is, when young children are asked a
metacognitive knowledge question like ‘Do you
know what is in the box?’, they just check whether
they can easily think of some plausible object name
and if so they have a sense of knowing and answer
affirmatively (‘Yes, I know’) to the knowledge
Kloo and Rohwer, 2012, 171
Can MM be overruled by AM?
• Noetic feelings can be overruled by additional
evidence, contrary feedback from others, repeated
failures in the task, etc.
• However: they cannot be suppressed (automatic,
• As a consequence of their representational format,
they are not open to:
– Inference
– revision
A two-system view of metacognition?
A contrast between norms.
• Fluency is a norm that is inherent to processing,
and that gives rise to specific feelings
(familiarity, confidence in perception or in
• Other norms, such as truth or plausibility, are
inherent the cognitive content to be accepted.
 Contrast between two forms of metacognitive
norms: experience-based and analytic.
The 2-system view revised
• Granting that System 1 generates
nonconceptual contents in a featural format,
the contrast with System 2 is one between
two ways of forming and using
System 1
System 2
• Vehicle-based
 Content-based
• Inflexible
 Flexible
• Economical
 Costly
• Nonconceptual
 Conceptual
• Gradient structure
 Componential structure
• Modular
 Non-modular
• Non inferential
 Inferential
• Inflexibility has nothing to do with the fact that
feelings are « generated by subpersonal
processes ». All our flexible thoughts are also
generated subpersonally.
• System1 inflexibility derives, rather, from the
nonconceptual format of representation that is
used to drive decision.
What kind of binding is there between
S1 and S2?
• The binding between the two systems is the
same as that studied in the philosophy of
perception between nonconceptual
protopropositional content, and propositional
What kind of binding is there between
S1 and S2?
• The nonconceptual content of perception is
inserted within a propositional format
including terms for concepts and objects.
• Analogously, children's NFs are redescribed in
conceptual terms.
What kind of binding is there between
S1 and S2?
• When a System 2 is present, agents have access
to propositional representations of their
cognitive goals, and can assess their cognitive
resources under new types of norms.
• Although this assessment may take marginal
advantage of NFs (e.g.: intelligibility), it is not
mainly based on cognitive emotions.
This presentation is available for download on :
Analytic norms