Chapter 4: Object Recognition
• What do various disorders of shape recognition tell us
•
•
about object recognition?
- Apperceptive visual agnosia (Ch. 2)
- Associative visual agnosia
- Perceptual categorization deficit
What do neuroimaging studies tell us about object
recognition?
The computational interpretation: What does cognitive
evidence suggest about constraints on the nature of the
neural representations underlying object recognition?
Object Recognition
• Dr. Harley has filled
you in on the
neuroscience side of
the miracle…
Farah (2000) – Fig. 4.1
Visual Agnosias
• Visual Agnosia: A blanket
term for impaired visual object
recognition following brain
damage when elementary
visual functions (acuity, visual
fields) are adequate.
• Apperceptive visual agnosia:
Inability to group the local
features into a coherent
perceptual representation (Ch
2)
• Associative visual agnosia:
Inability to recognize visually
presented objects, despite
having a coherent perceptual
representation
Visual Agnosias: Copying and
drawing from memory.
Apperceptive agnosia
Associative agnosia
Copying
Copying
Memory
Associative visual agnosia: Criteria
• Patients can form percepts (do perceptual
•
•
•
grouping) unlike apperceptive agnosia patients.
They can see an object well enough to describe
its appearance, to draw it, or to succeed in a
same/different test of appearances.
They have difficulty recognizing visually
presented objects (can’t name or sort objects by
category)
They can demonstrate knowledge of objects
from other sensory modalities
Associative agnosia behavior: An
impairment of shape perception
• They can copy complex objects.
BUT their copying behavior is
abnormal – HJA took 6 hours to do
the cathedral.
• They are very sensitive to visual
quality of stimuli. They recognize
real objects better than line
drawings; face recognition of
unfamiliar people is impaired by
changing lighting conditions.
• They make visual shape recognition
errors; they might call a baseball
bat, a paddle, knife, baster, etc.
Associative Visual Agnosia
• The human analog of the IT-lesioned
monkeys.
– They fail to recognize objects because they
fail to represent shape normally.
• Slow, slavish copying
• Sensitivity to visual quality of the stimuli
• Visual shape errors
Regions of brain damage associated
with different types of visual agnosia
• Apperceptive Agnosia:
•
Diffuse damage to the
occipital lobe and
surrounding areas
Associative Agnosia:
Occipitotemporal
regions of both
hemispheres
From Banich (2004)
Perceptual categorization deficit
• Difficulty recognizing
•
objects viewed from
unusual perspectives or
uneven illumination
conditions (Warrington &
Taylor, 1973)
Initially assumed to
reflect an impairment of
viewpoint-invariant object
recognition
BUT: Is Perceptual Categorization
Deficit really about loss of shape
representations?
• Not impaired in the real world.
• Not impaired under all viewing conditions per se.
• Mainly impaired when matching an USUAL to an
•
UNUSUAL view; normal people have similar, less
serious, type of difficulty.
Associated with unilateral RH lesions, usually in
parietal cortex, not inferotemporal cortex
Perceptual categorization deficit
• Demonstrates value of examining evidence
carefully and trying to link it with what is
known to see whether patterns fit or not.
• In this case, the pattern differs greatly
from other visual agnosias.
Functional Neuroimaging Studies:
• Goal of PET, fMRI studies: To localize the
•
psychological process(es) of interest
Research design: Measure brain activity in at
least two conditions: a control (baseline)
condition and an experimental condition.
– Psychological process localized by subtracting brain
activity in the control (baseline) condition from brain
activity in the experimental condition.
PET Imaging (Posner & Raichle, 1994)
Upper row: Control PET scan
(resting while looking at static
fixation point) is subtracted from
looking at a flickering
checkerboard stimulus positioned
5.5° from fixation point.
Middle row: Subtraction
procedure produces a somewhat
different image for each of 5
subjects.
Bottom row: The 5 images are
averaged to eliminate noise,
producing the image at the bottom.
Where in the brain does visual
recognition occur?
• Visual recognition
•
•
localized to the
posterior half of the
brain – really?
Whoops!!! Why did
these studies fail to
localize visual
recognition?
Poor methodology!
Neuroimaging methods: The
subtraction technique
• Using the subtraction technique effectively
requires that the baseline and experimental
conditions differ only in the process of interest.
– Requires careful logical analysis of the task to
determine the best comparison conditions.
• If there are many differences between the
baseline and experimental conditions, it is
difficult to interpret the results of the study.
Neuroimagining Studies of Object
Recognition: Comparing baseline and
experimental conditions
Sequence of trial events
– Fixation point
– Stimulus presentation
– Task (two types)
Passive Viewing Task:
Are stimuli comparable
(e.g., size, complexity)
Mind is not a vacuum
Active Viewing Task:
What is task?
What is response?
Sergent et al (1992a):
Active viewing task.
• Baseline condition
- View fixation point
• Experimental
condition
- View fixation point
- See line drawing of
object
- Decide whether the
object is living or nonliving
- Make Yes/No
response
Localizing Visual Recognition
• Is human visual recognition supported by
one general purpose recognition system or
are there specialized modules for
recognizing objects, faces, printed words?
• Good functional neuroimaging studies do
exist to test these possibilities and will be
discussed in Chapters 5 & 6.
Constraints on the nature of shape
representations in IT: Coordinate systems
• Cannot be simple viewer-centered or
environmentally-centered
– IT cells respond to a given shape over
changes in the position, size, and picture
plane orientation of object
– Impairments of IT lesioned monkeys:
suggest they have lost abstract shape
representation
Constraints: Coordinate systems cont.
• Two possibilities:
1. Object-centered
coordinate system
2. A cluster of multiple
viewer-centered object
representations plus
the ability to transform
one representation to
another as necessary
(a la Multiple Views
model of Tarr, 1995).
Constraints: Coordinate systems cont.
• Farah prefers the Multiple Viewer-Centered
representation option because:
– Position, size, and orientation invariance shown by IT
cells is not perfect; consistent with some views being
better learned than others.
– IT neurons can selectively learn arbitrary associations
between pairs of stimuli, a prerequisite for deriving
invariance from viewer-centered representations.
– Consequently, because cell activity takes time to
decay and because seeing different views of an object
tends to be clustered in time, the correlation
(association) between several different retinotopic
views of the same object could be learned.
Constraints:
Primitives and Organization
• Primitives
– Cannot be contours
– Could be either surface-based (2D) or
volume-based (e.g., 3D geon-like parts).
• Organization
– Hierarchical? We really don’t know much
about how multipart objects are represented
Constraints: Implementation
Symbolic Model vs Neural net
• Is a perceptual representation created and then
compared to a memory representation? (Implies
that the memory representation can be destroyed,
yet the perceptual representation remains intact).
Unlikely, since no such evidence exists.
• Does the input get coded and recoded in a
succession of neural nets as it is processed through
the visual system? (Implies that impairment in test
of object memory is associated with perceptual
impairment). Yes, this seems true in associative
agnosia.
Constraints: Implementation
• Distributed
representation more
likely than local
representation
– Single unit recording
– Graceful degradation
following damage to
IT