Computational Maps in the Visual...
Transcript of Computational Maps in the Visual...
Computational Maps in the Visual Cortex
Risto Miikkulainen
Department of Computer Sciences
and Institute for Neuroscience
The University of Texas at Austin
Joint work with Jim Bednar, Yoonsuck Choe, and Joseph Sirosh
Supported in part by NIMH 1R01-MH66991, NSF IIS-9811478 & IRI-9504317
Understanding Vision
(Logothetis 2002)
• How is a system as complex as the human visual system constructed?
• How can it be both genetically and environmentally determined?
• How does its structure support functions such as perceptual grouping?
Role of Computational Modeling
V1
LGN
ON OFF
Retina
• Computational model is an artificial subject with full access– Test hypotheses computationally, make predictions
• Computational theory of the visual cortex– Build better artificial systems– Improve medical treatment
Human Visual System
chiasmOptic
Right
Left
Left LGN
Right LGN
Visual field
Left eye
Right eye
(V1)cortexvisualPrimary
• Retina, LGN, V1...etc.
• Structure well known
Receptive Fields
Retina
ON/OFF
OFF/ON
LGN
ON/OFF
OFF/ON
V1
Simple 2-lobe
Simple 3-lobe
Spatiotemporal
• Center-surround; static and moving lines; combinations
Columnar Organization of V1
(Kandel et al. 1991)
• Roughly hierarchical ordering:– Retinotopy, OD, OR, DR– Color, spatial frequency, disparity?
• Within column, similar responses: 2D structure
Measuring Cortical Maps
• Surface reflectance changes with activity
• Optical imaging can be used to detect
Orientation Map
(7.5 mm × 5.5 mm in macaque V1; Blasdel, 1992)
• Preferences mapped systematically
• Linear zones, pinwheels, saddles, fractures
Orientation & Ocular Dominance Map
(4 mm × 3 mm in macaque V1; Blasdel, 1992)
• Systematic interactions
– OD, OR boundaries at right angles
– Pinwheels, saddles in the middle
Orientation & Direction Map
(1.4 mm × 1.1 mm in ferret V1; Weliky et al. 1996)
• Systematic interactions
– OD, OR boundaries at right angles
– OR patches contain opposite DR
Lateral Connections
(2.5 mm × 2 mm in tree shrew V1; Bosking et al. 1997)
• Link to similar responses
• Patchy structure, extend along OR preference
Development
(4 mm × 3 mm OR+select in ferret V1; Chapman et al. 1996)
• Structure emerges during development
• Some prenatally, much postnatally
• How and why?
LISSOM Model
Right retinaLeft retina
3
2
1
0
OFFONOFFON
LGN
V1
• Combined OR, OD, DR
• Retina, LGN, V1 (+ other areas)
• 2D sheets, afferent and lateral connections
• Hebbian learning in V1
Activation
Retinal activation LGN response V1 initial V1 settled
• Luminance adjustment in retina
• Sharpening in LGN (ON−OFF shown)
• Settling in V1:
η′i = σ (ΣkχkAki +ΣjηjEji − ΣjηjIji)
Adaptation
Itera
tion
0Ite
ratio
n10
,000
Afferent Lateral excitatory Lateral inhibitory
• Normalized Hebbian learning: A′ki =
Aki+αχkηi
Σmn(Aki+αχkηi)
→ Input-driven self-organization
• Pruning unused connections
• Results in realistic receptive fields, patchy lateral connections
Orientation Map
OR+selectivity, iteration 0 OR+selectivity, iteration 10,000
• Systematic preferences emerge
• Similar structures as in biology
Orientation & Ocular Dominance Orientation & Direction
OR & OD & DR Map Lateral Connections
Wei
ghts
His
tC
onne
ctiv
ity
Iso-OR patch Pinwheel Saddle Fracture
• Link similar responses• OR primary factor• Matches biology; detailed predictions
Self-Organization Conclusions
• How is V1 constructed?
– Input-driven self-organization
• Predictions:
– Input deprivation (e.g. strabismus)
– Connection patterns
– Plasticity
– Illusions and aftereffects
– Visual coding
What Is the Goal of Visual Coding?
V1
LGN
ON OFF
Retina
• Representing the important features of the input
• Efficient use of resources:
Can represent more information within a limited system
How is Such a Coding Constructed?
Initial response Redundancy-reduced
sparse response
• Not by reducing units: V1 is much larger than the retina
• Could be a sparse code with few active units
• Need to make sparse by reducing redundancy(Barlow 1972; Atick 1992; Field 1994; Simoncelli & Olshausen 2001)
Lateral Connection Hypothesis
(2.5 mm × 2 mm in tree shrew V1; Bosking et al. 1997)
• Afferent connections respond to input features
• Inhibitory lateral connections decorrelate the response
– Connect neurons that respond to similar inputs
– Response of one neuron can be predicted from the other
– Can be suppressed without losing information
Testing the Hypothesis
V1
LGN
ON OFF
Retina
• Difficult to test experimentally
– Requires many neurons, short time scales
• Can be tested in computational models
Does LISSOM Form a Sparse Code?
Retinal activation Initial V1:
Kurtosis 37.9
Settled V1:
Kurtosis 63.0
• Self-organize a LISSOM map
• Measure kurtosis of the response
• → The settled response is sparser
Does LISSOM Reduce Redundancy?
Retinal activation Reconstruction
from initial V1:
Avg. RMS error 0.094
Reconstruction
from settled V1:
Avg. RMS error 0.094
• Reconstruct the input from V1 activity
• Nonlinear: train a backprop net to map back
• → No information lost
Is Self-Organization Necessary?
Retinal activation Settled V1:
Kurtosis 63.0
SoG-settled V1 :
Kurtosis 60.1
• Isotropic (Sum-of-Gaussians; SoG) lateral
connections instead
• Can be adjusted to match kurtosis
• → Sparse code can be formed
Is Self-Organization Necessary?
Retinal activation Reconstruction
from LISSOM V1:
Avg. RMS error 0.094
Reconstruction
from SoG V1:
Avg. RMS error 0.137
• Reconstruction no longer works!
• Information reduced, not just redundancy
• → Self-organization is necessary
• → Forms a sparse, redundancy-reduced code
Nature vs. Nurture
(Johnson and Morton 1991)
• Development through input-driven self-organization
• But some order appears innate
– E.g. orientation maps
– E.g. newborn face preferences
Newborn Face Preferences
0o
10o
20o
30o
40o
50o
60o
Fin
al tr
acki
ng o
rient
atio
nF
Within 1 hour
0o
10o
20o
30o
40o
50o
60o
Fin
al tr
acki
ng o
rient
atio
nF
21 hours(Johnson et al. 1991)
• Significant preference for face-like schematics
• Genome too small to specify connectivity, behavior
• Three-dot patterns strongest; why?
Retinal Waves
(1 mm × 1 mm in ferret retina; Feller et al. 1996)
• Traveling waves in the retina before birth
• Could serve as input for self-organization
PGO Waves
V1
LGN
(Marks et al. 1995)
• Ponto-geniculo-occipital waves
• Shape unknown, but activates V1
• Could introduce the three-dot bias
HLISSOM Model
LGN
ON OFF
V1
Face−selective area (FSA)
Retina PGOpattern
generator
• Include PGO & FSA
sheets
• Three-dot input patterns
in PGO
• Study prenatal and
postnatal
self-organization
Newborn LISSOM Face Preferences
Ret
ina
LGN
V1
FS
A
2728
0.0
927
5.8
836
5.5
664
4.0
1094
0.0
994
0.0
931
0.0
807
0.0
275
0.0
• Matches newborn preferences in every known case
Newborn LISSOM Face Preferences (2)
Ret
ina
LGN
V1
FS
A
4756
11.7
5778
7.5
6932
6.1
4797
11.1
4796
0.0
7614
0.0
7858
0.0
5072
0.0
4845
18.1
• Prefers top-lit faces; not objects
• Images not tested on infants
Effect of Pattern Types
PG
OLG
NF
SA
-RF
Sch
Img
1.0
1.0
0.8
1.0
0.8
0.8
0.3
0.9
0.5
0.4
• Three dots not the only possible pattern
• Not all patterns work
Pattern Generation Conclusions
• How are nature and nurture combined?
– Through internal pattern generation
• Predictions
– Types of internal patterns
– Postnatal decline of preferences
– Holistic perception of the face develops
– Mother preferences develop
Perceptual Grouping
Proximity Good continuation World knowledge
• Perceiving whole objects
• Low-level based on “Gestalt” principles
• Mediated by lateral connections in V1?
PGLISSOM Model
ON
V1
Retina
GMAP
SMAP
• Self-organization needs long-range
inhibition
• Grouping needs long-range
excitation
• → 2-layer model of the column
Leaky Integrator Neuron
χ~
η~
η~
ζ~
Inhibitory lateral connections
Excitatory lateral connections γ
γ
γ
γ
A
E
I
C
v S
A
σ
a
r
E
I
b
γ
C
θ
Σ
Σ Σ
Spike generator
Intra−columnar connections
Afferent connections
Σ
Σ
Σ
• Binding and segmentation by synchronization
• Need spiking neurons
Self-Organized Lateral Connections
100
10
1
0.1
0.01
δ
Likelihood ratio
= 1.23o
φ = 90o
φ = 0o
0.00010.001
0.010.1
1
δ
Relative probability
= 0 o
= 90 o
φ
= 27
φ
Edge cooccurrence in nature(Geisler et al. 2001)
Lateral excitation in GMAP
• PGLISSOM self-organizes like LISSOM
• Lateral connections match visual environment
Contour Integration Process
1
2
36
54
7
8
9
1
2
3�
4
5�
6
7
8�
9�
0�
100 200�
300�
400�
500�
0◦ of orientation jitter
21
3
5 6
98
7
4 1
2
3�
4
5�
6
7
8�
9�
0�
100 200�
300�
400�
500�
50◦ of orientation jitter
• Synchronizes continuous contours
• Depends on how “good” the contour is
PGLISSOM vs. Human Performance
0o
10o
20o
30o
40o
50o
60o
70o
60%
70%
80%
90%
100%
Hum
an: A
ccur
acy
(por
tion
corr
ect)
Subject WSG
0o
10o
20o
30o
40o
50o
60o
70o
Orientation jitter
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Mod
el: c
orre
latio
n (r
)
Model
• Depends on jitter like human performance
Contour Segmentation
12
3
4
5
6
7
89
1
2
3�
4
5�
6
7
8�
9�
0�
100 200�
300�
400�
500�
• Multiple contours by alternating
• Upto 5-9 contours
Contour Completion
123
45
1
2�
3�
4�
5�
0�
100 200 300�
400 500�
• Filling in gaps
• Basis for edge-induced illusory contours?
Illusory Contours
Open1
23 4
56789
10
1112 15
13
14
123
�45
�678
�9
�101112131415
0�
100 200 300�
400 500�
Closed1
23 4
56789
10
1112 15
13
14
123
�45
�678
�9
�101112131415
0�
100 200 300�
400 500�
• Kanizsa: proximity & continuation
• Closed contours easier
• Matches human performance
Perceptual Grouping Conclusions
• How does the structure support functions like grouping?
– Synchronization mediated by self-organized lateral connections
• Predictions:
• Effect of activation decay, noise, refractory period on synchronization
• Image statistics → lateral connectivity → performance
– Frequency, curvature, etc. differ across visual fields
– Performance differs in fovea vs. periphery, upper vs. lower hemifield
Future Work
• Self-organization
– Color, frequency, disparity
– Hierarchy, feedback, multimodal integration
• Development
– Characterizing internal patterns
– Constructing complex systems
• Grouping
– Verify synchronization hypothesis with TMS
– Line-end-induced illusions in V2?
Topographica
• General simulator for cortical maps (v0.8.2 Feb 2006)
Conclusion
• Wealth of data + powerful computing available
• Neuroscience research in vitro, in vivo, in silico
• Computational theory of the visual cortex
– Continuously adapting self-organizing system
– Shaped by internal and external input
– Lateral connections play a major role
• Exciting possibilities for future work
Further Details
(Springer, 2005)
Demos, software, etc.:www.computationalmaps.org