Biases: An Example
Non-accidental properties: Properties that appear in an image
that are very unlikely to have been produced by chance, and
therefore are likely to reflect properties of the 3-D world.
view-point invariant properties
straight lines
parallel lines
mental representation
of objects
book
cat
key
tree
etc..
view-point invariant
?
mental representation
of bars of light
view-point dependent
mental representation
of objects
book
cat
key
tree
etc..
view-point invariant
1) Direct models
mental representation
of bars of light
view-point dependent
mental representation
of objects
book
cat
1) Direct models:
Lades et al Model
Poggio & Edelman Model
mental representation
of bars of light
key
tree
etc..
view-point invariant
•Input layer like representation in V1
(called “Gabor Jets”)
•Inputs on that layer are matched to
inputs in a memory layer
•The object is identified based on
the match with least distortion
•Input layer like representation in V1
•3-layer network is trained to rotate
all views of an image to one view.
•The hidden units are seen as a way
of rotating images to match memory
images (“radial basis functions”)
view-point dependent
mental representation
of objects
book
cat
1) Direct models:
Lades et al Model
Poggio & Edelman Model
mental representation
of bars of light
key
tree
etc..
view-point invariant
•Input layer like representation in V1
(called “Gabor Jets”)
•Inputs on that layer are matched to
inputs in a memory layer
•The object is identified based on
the match with least distortion
•Input layer like representation in V1
•3-layer network is trained to rotate
all views of an image to one view.
•The hidden units are seen as a way
of rotating images to match memory
images (“radial basis functions”)
view-point dependent
mental representation
of objects
book
cat
key
tree
etc..
view-point invariant
2) View-point invariant
mental representation
of non-accidental
properties of an image
Lowe’s SCERPO Model
Ullman’s Model
mental representation
of bars of light
view-point invariant
•Input layer takes information
represented as it is in V1
•view-point invarient information
is extracted
•this allows the input image to be
rotated in order to fit an image
stored in memory
view-point dependent
The problem with all of these theories:
The representation of objects in memory is stored as
a two-dimensional image, which visual images are
rotated, distorted, and matched to.
But in actuality, objects are three dimensional things
in the world.
So lets make a model which has the basic units that
make up mental representations of objects being
three-dimensional solids, rather than lines and edges.
mental representation
of objects
mental representation
of geons
book
cat
key
tree
etc..
view-point invariant
view-point invariant
mental representation
of non-accidental
properties of an image
view-point invariant
mental representation
of bars of light
view-point dependent
Geons
straight, parallel
curved, parallel
corner
Y intersections
Y intersections
mental representation
of objects
mental representation
of geons
book
cat
key
tree
etc..
view-point invariant
view-point invariant
mental representation
of non-accidental
properties of an image
view-point invariant
mental representation
of bars of light
view-point dependent
mental representation
of objects
book
cat
key
briefcase
etc..
mental representation
of geons
view-point invariant
what processing
rectangle unit
cylinder unit
cone unit
tube unit
view-point invariant
where processing
above unit
below unit
left of unit
righ of unit
temporal binding
Evidence for geons?
Experiment 1: Visual Priming
Reaction Time: 900 ms.
(response)
time
(response)
Evidence for geons?
Experiment 1: Visual Priming
Reaction Time: 700 ms.
(response)
(response)
time
Evidence for geons?
Experiment 1: Visual Priming
Reaction Time: 800 ms.
(response)
(response)
time
Evidence For Geons: Priming Studies
First
Second
Shared?
Lines
Edges
Geons
Basic
Cat.
Response
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
Time
700 ms
700 ms
No
No
No
No
900 ms
Evidence For Geons: Priming Studies
First
Second
Shared?
Lines
Edges
Geons
Spec.
Cat.
Basic
Cat.
Resp.
Yes
Yes
Yes
Yes
Yes
Time
700 ms
No
No
Yes
Yes
Yes
800 ms
No
No
No
Yes
Yes
800 ms
Objects, Faces and Rotation:
Objects
1020
1000
980
960
940
920
900
0
60
120
180
240
300
Face Recognition
Recognizing faces is an entirely different problem
(computationally) from recognizing objects?
Objects
• “M” shaped function
Faces
• Shape of rotation function?
Face Recognition Experiment
Objects, Faces and Rotation:
Faces
1060
1040
1020
1000
980
960
940
920
900
0
60
120
180
240
300
Face Recognition
Recognizing faces is an entirely different problem
(computationally) from recognizing objects?
Objects
Faces
• “M” shaped function
• Less affected by illumination
• Shape of rotation function?
• More affected by illumination
Recognition by components
(Geon Theory)
Recognition by coordinates
(Templates)
different processes?
Double Dissociation
Patient Studies
• Prosopagnosia: Objects can be recognized, Faces can not.
Processes responsible for
object recognition:
Processes responsible for
face recognition:
Double Dissociation
Patient Studies
• Prosopagnosia: Objects can be recognized, Faces can not.
• (Patient CK): Faces can be recognized, Objects can not
Processes responsible for
object recognition:
Processes responsible for
face recognition:
Double Dissociation
Patient Studies
• Prosopagnosia: Objects can be recognized, Faces can not.
• (Patient CK): Faces can be recognized, Objects can not
Neuroimaging Studies
• Faces activate the FFA
area of the temporal lobe
• Objects activate the PPA
area of the temporal lobe
Double Dissociation
General Form
IF:
The ability to perform some task X is affected by or correlated
with some factor A (damage to an area, activity in a part of the
brain, or performing some other task) but not factor B (damage
to a different area, activity in a different part, or some other task),
AND:
The ability to perform some task Y is affected by or correlated
with some factor B but not factor A,
THEN:
There is a double dissociation between X and Y, and the
mechanisms required to perform them are functionally independent.
Dissociation
What and Where Pathways
Some patients
Can
identify objects by shape
Cannot
Say where objects are
Cannot navigate the world
Damage
Parietal lobe
Dorsal visual stream
X
Dissociation
What and Where Pathways
Some patients
Cannot
identify objects by shape
Can
Say where objects are
Cannot navigate the world
Damage
Temporal lobe
Ventral visual stream
X
Dissociation
What and Where Pathways
Patient D.F.
Cannot
Explicitly line up a line with a line
Task: Take an envelope and line it up with this
line
Can
Use motor system (part of “where” pathway)
to do same task just failed at
Task: Take this envelope and “mail the letter”
pretending that this line is a mailbox
Dissociation
Two face recognition systems
• Recognize faces normal way “ Hey I know you”
• Recognize by change in skin conductance for
familiar faces
Dissociation
Two face recognition systems
• Person A can recognize normal way but no
change in skin conductance
• Person B cannot recognize in normal way but
does have change in skin conductance