261.21 neocortical layer 4 is a pluripotent function
TRANSCRIPT
261.21 Neocortical Layer 4 is a Pluripotent Function LinearizerOleg V. Favorov and Olcay Kursun
Department of Biomedical Engineering, University of North Carolina School of Medicine, Chapel Hill, NC 27599
INTRODUCTION
In Machine Learning/Pattern Recognition, a highly
effective kernel-based strategy for dealing with nonlinear
problems is to transform the input space into a new
“feature” space, in which the problem becomes linear and
more readily solvable with efficient linear techniques. We
propose that a similar “problem-linearization” strategy
might be used by neocortex. A mathematically abstract
elaboration of such a problem-linearization strategy
produces a computational system that closely resembles
the real cortical layer 4 in its structural and functional
properties. We demonstrate this close match between
theoretical and real cortical properties on layer 4 of the cat
primary visual cortex.
Problem-linearization
strategy of
transforming the input
space into a “feature”
space
Each cortical area computes
certain nonlinear functions F
over its afferent inputs.
Each cortical area might benefit
from a problem-linearization
strategy in learning its functions
HYPOTHESIS: Layer 4 implements a function-
linearization strategy for the upper layers 2/3
Layer 4 transform must be “blind,”
without feedback from layers 2/3.
Layer 4 transform must be
“pluripotent” – optimized to
make linear as broad a repertoire
of potential functions as possible.
IMPLEMENTATION OF FUNCTION-
LINEARIZATION STRATEGY UNDER NEURAL
CONSTRAINTS
Anti-Hebbian lateral connections are needed to drive
neighboring Layer 4 neurons to diversify their preferred
directions in the stimulus space by modifying their Hebbian
afferent connections.
BIOLOGICAL INTERPRETATION OF THE MODEL
Experimental evidence:
Egger V, Feldmeyer D, Sakmann B (1999)
Coincidence detection and changes of
synaptic efficacy in spiny stellate neurons in
rat barrel cortex. Nature Neuroscience 2:
1098-1105.
TEST OF FUNCTION-LINEARIZATION HYPOTHESIS ON VISUAL INPUTS (comparison with Layer 4 of cat V1)
Natural images were used to develop plastic connections in
the model.
THALAMIC (LGN) MODEL
LAYER 4 MODEL
PLURIPOTENCY TEST
Pluripotency of L4 transform – capacity to represent linearly
any arbitrary nonlinear function of the afferent input patterns.
LGN connections
of all L4 cells
EMERGENT LGN CONNECTIONS AND RECEPTIVE FIELDS
End-stopped RFs:
15% (25%)
Number of RF
subfields:
1-4 (1-5)
Av. number of RF
subfields:
2.7 (2.45-2.65)
RF aspect ratio:
3.8 (4.3-4.5)
ORIENTATION TUNING (OT)
- simple-cell type of response to gratings
- HWHH of OT: 15o
(16o)
- OT is contrast-invariant
- average optimal spatial frequency: 1o
(0.86o)
- OT is tighter for finer gratings
Equation
offers a close mathematical representation of the structure and function
of local layer 4 domains
CONCLUSION
Hebbian Feed- Recurrent Anti-Hebbian
thalamic forward inhibition recurrent
input inhibition excitation
OT of output is close to OT of LGN input -