Lecture 6: Object perception

PSYC102
10/22/2019
 Overview

Dorsal and ventral visual streams

Neural mechanisms of object recognition, ventral

visual stream

2
Beyond V1: Dorsal and Ventral pathways

Parietal lobe

Temporal lobe
3
Parvocellular layers
(3-6) Functional
Specialization
Magnocellular layers
(1 & 2)

And low-light vision

4
Dorsal and Ventral visual pathways: classic account

5
 Lesion or Ablation Experiments to assess function of visual
pathways

• First, an animal is trained to indicate perceptual

capacities
• Second, a specific part of the brain is removed or
destroyed (‘ablated’)
• Third, the animal is retrained to determine which
perceptual abilities remain
• The results reveal which portions of the brain are
responsible for specific behaviors

6
 What and Where Pathways
• Ungerleider and Mishkin (circa 1980)

– Object discrimination
problem
• Monkey is shown an object
• Then presented with two choice
task
• Reward given for detecting the
target object

– Spatial landmark
discrimination problem
• Monkey is trained to pick the
food well next to a cylinder

7
 What and Where Pathways

• Ungerleider and Mishkin (cont.)

– Using ablation, part of the

parietal lobe was removed
from half the monkeys and
part of the temporal lobe
was removed from the
other half
– Retesting the monkeys
showed that:
• Removal of temporal lobe tissue
resulted in problems with the object
discrimination task - What pathway
• Removal of parietal lobe tissue resulted
in problems with the landmark
discrimination task - Where pathway

8
 Dorsal and Ventral pathways:
a second interpretation
Milner and Goodale’s idea:
• Dorsal pathway (parietal cortex): not where, but ‘how’

9
 What and How Pathways –
Neuropsycholgical Evidence

• Behavior of patient D.F.

– Damage to ventral pathway due to carbon
monoxide intoxication in 1988

10
• Behavior of patient D.F.
– Damage to ventral pathway due to gas leak

D.F. can’t tell you the orientation of the slot, but can
actively post a letter into it
Ventral stream = what, Dorsal stream = How. 11
What vs how (ventral damage)
High level summary

13
OBJECT RECOGNITION
 Inverse projection problem: An image
on the retina can be caused by an
infinite number of objects.

Fundamentally ambiguous at the level of

the retinal image…need to impose
additional constraints
Object recognition…very hard problem
 The Gestalt Approach
• Stimulated the founding of Gestalt psychology in the
1920s by Wertheimer, Koffka, and Kohler
• According to Gestalt psychologists, the whole is
different than the sum of its parts. Gestalt is a
German word meaning configuration or pattern.
• Perception is not built up from sensations but is a
result of your brain imposing perceptual
organization on incoming stimuli
• Gestalt principles do not make strong enough
predictions to qualify as “laws”
– They are better thought of as heuristics - “best
guess rules”
 Gestalt Organizing Principles - continued

• Good continuation – continuous shapes

viewed as single segmented objects
• Proximity/Similarity - things that are near to
each other are grouped together
• Common fate - things moving in same
direction are grouped together
• Common region - elements in the same
region tend to be grouped together
• Uniform connectedness - connected region of
visual properties are perceived as single unit
 The Gestalt laws of perceptual organization

Good continuation – helps us perceive this as one continuous object, not a

bunch of disconnected overlapping objects
On (Google) Earth
• http://www.richardgregory.org/experiments/index.htm
 Simple computational model:
Recognition by Components (RBC)
Biederman (1985)

Geons (“Geometric Ions”)

Each geon is uniquely identifiable from most
viewpoints (viewpoint invariant).
Only 36 geons needed to make thousands of
objects.
Objects can be identified if the geons can be
identified:
which geons are present?
what is the spatial relation among geons?
 Structural-Description Models

Recognition by Components (RBC)

 Recognition by Components
• Strengths
– Viewpoint invariant
– Represent 3-D structure
• Weaknesses
– Complexity of representation
– Doesn’t easily represent subtle metric
differences (e.g., distance between eyes)
– Recognition is at the level of categories
(chair vs. table) rather than individuals (my
office chair vs. my kitchen chair)
Deep Neural Networks (DNNs)
~17,200 papers in the last year!
 Summary
• Brain uses basic Gestalt grouping principles to
organize input

• Computational theories like RBC can help to

explain how we build complex objects based on a
simpler set of components (but its not a
complete theory!)

• DNNs are much more nuanced, but lack some

sophisticated heuristics…
30
How Does the Brain Process Information About Objects?
Visual input is extremely variable – somehow need position/rotation invariance (or
tolerance)
 View invariant responses to familiar objects in monkey area IT

The neuron’s responses are similar within each object across viewpoints.
Booth and Rolls, 1998
Viewpoint and size invariant responses – just
signal the feature that you are tuned to
40
 Single unit responses in area TE – very stable across time to learned images!

Bondar IV, Leopold DA, Richmond BJ, Victor JD, et al. (2009) Long-Term Stability of Visual Pattern Selective Responses of Monkey Temporal Lobe
Neurons. PLoS ONE 4(12): e8222. doi:10.1371/journal.pone.0008222
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0008222
 Despite stability, responses are flexible during learning…
This balance between stability and flexibility during learning is key…

Series of familiar objects Series of novel objects

Meyer, Walker, Cho, Olson, Nature Neuroscience, 2014

 Large scale organization of ventral
pathway

• Equally responsive to all features/objects?

• No…seems to have specialized chunks of
cortex that perform somewhat distinct
perceptual functions
• Often referred to as modules

44
 ‘Modules’ in the ventral pathway of the macaque and humans

Nancy Kanwisher, MIT

Figure 4.18 (a) Monkey brain showing location of the inferotemporal cortex (IT) in the lower part of the temporal lobe. (b) Human brain
showing location of the fusiform face area (FFA) in the fusiform gyrus, which is located under the temporal lobe.
45
 FFA/PPA/LOC – innate?

Kalinit Grill-Spector,
Stanford

Fusiform Face Area (FFA): Faces > objects/scenes

Parahippocampal Place Area (PPA): Scenes/Landmarks > Faces/Objects

Lateral Occipital Complex: Intact Objects > Scrambled (nonsense) objects

 Not quite so simple…
• Actually not just a single ‘module’
• Face identity likely represented by distributed
code
 Face patches, Doris Tsao, CalTech

MacArthur "Genius” Award!!!

Ethan M. Meyers, Mia Borzello, Winrich A. Freiwald and Doris Tsao, 2015, Journal of
Neuroscience
 All patches discriminate individual
faces and objects to varying degrees

Ethan M. Meyers, Mia Borzello, Winrich A. Freiwald and Doris Tsao, 2015, Journal of
Neuroscience
 Summary object recognition
• Brain uses basic Gestalt grouping principles to organize
input

• Computational theories like RBC can help to explain how

we build complex objects based on a simpler set of
components (but not so good at fine details)

• Functional specialization, but not just one ”module” for

each perceptual domain (e.g. faces, houses, etc)

• Neural activity in ventral visual cortex corresponds very

closely to your subjective perceptual experience

56