Overview paper and theses on the RF-LISSOM project

[jbednar@cs.utexas.edu]

The following paper (to appear in Psychology of Learning and Motivation)
gives an overview of the RF-LISSOM project on modeling the primary
visual cortex, ongoing at the UTCS Neural Networks Research Group.

For those seeking details, please refer to the dissertation and the
theses also announced below. All these publications (and others) are
available from our web page at http://www.cs.utexas.edu/users/nn (under
publications, self-organization; the direct ftp addresses are included
below as well).  Public domain LISSOM code, for self-organization in
laterally connected maps, will be announced in the near future.

-- The Authors

-------------------------------------------------------------------------
SELF-ORGANIZATION, PLASTICITY, AND LOW-LEVEL VISUAL PHENOMENA 
IN A LATERALLY CONNECTED MAP MODEL OF THE PRIMARY VISUAL CORTEX

Risto Miikkulainen, James A. Bednar, Yoonsuck Choe, and Joseph Sirosh
Department of Computer Sciences, The University of Texas at Austin.
In R. L. Goldstone, P. G. Schyns, and D. L. Medin (eds.), Psychology
of Learning and Motivation, vol. 36, 1997 in press (36 pages).
ftp://ftp.cs.utexas.edu/pub/neural-nets/papers/miikkulainen.visual-cortex.ps.Z

Based on a Hebbian adaptation process, the afferent and lateral
connections in the RF-LISSOM model organize simultaneously and
cooperatively, and form structures such as those observed in the primary
visual cortex. The neurons in the model develop local receptive fields
that are organized into orientation, ocular dominance, and size
selectivity columns. At the same time, patterned lateral connections
form between neurons that follow the receptive field organization. This
structure is in a continuously-adapting dynamic equilibrium with the
external and intrinsic input, and can account for reorganization of the
adult cortex following retinal and cortical lesions. The same learning
processes may be responsible for a number of low-level functional
phenomena such as tilt aftereffects, and combined with the leaky
integrator model of the spiking neuron, for segmentation and
binding. The model can also be used to verify quantitatively the
hypothesis that the visual cortex forms a sparse, redundancy-reduced
encoding of the input, which allows it to process massive amounts of
visual information efficiently.

-------------------------------------------------------------------------
A SELF-ORGANIZING NEURAL NETWORK MODEL OF THE PRIMARY VISUAL CORTEX

Joseph Sirosh
Department of Computer Sciences, The University of Texas at Austin.
PhD Dissertation and Technical Report AI95-237, August 1995 (137 pages).
ftp://ftp.cs.utexas.edu/pub/neural-nets/papers/sirosh.diss.tar

This work is aimed at modeling and analyzing the computational processes
by which sensory information is learned and represented in the
brain. First, a general self-organizing neural network architecture that
forms efficient representations of visual inputs is presented.  Two
kinds of visual knowledge are stored in the cortical network:
information about the principal feature dimensions of the visual world
(such as line orientation and ocularity) is stored in the afferent
connections, and correlations between these features in the lateral
connections. During visual processing, the cortical network filters out
these correlations, generating a redundancy-reduced sparse coding of the
visual input.  Through massively parallel computational simulations,
this architecture is shown to give rise to structures similar to those
in the primary visual cortex, such as (1) receptive fields, (2)
topographic maps, (3) ocular dominance, orientation and size preference
columns, and (4) patterned lateral connections between neurons.  The
same computational process is shown to account for many of the dynamic
processes in the visual cortex, such as reorganization following retinal
and cortical lesions, and perceptual shifts following dynamic receptive
field changes.  These results suggest that a single self-organizing
process underlies development, plasticity and visual functions in the
primary visual cortex.

-------------------------------------------------------------------------
TILT AFTEREFFECTS IN A SELF-ORGANIZING MODEL OF THE PRIMARY VISUAL CORTEX

James A. Bednar
Department of Computer Sciences, The University of Texas at Austin.

Masters Thesis and Technical Report AI97-259, January 1997 (104 pages).
ftp://ftp.cs.utexas.edu/pub/neural-nets/papers/bednar.thesis.tar

The psychological phenomenon known as the tilt aftereffect was used to
demonstrate the functional properties of RF-LISSOM, a self-organizing
model of laterally connected orientation maps in the primary visual
cortex.  The same self-organizing processes that are responsible for
the development of the map and its lateral connections are shown to
result in tilt aftereffects as well.  The model allows analysis of
data that are difficult to measure in humans, thus providing a view of
the cortex that is otherwise not available.  The results give
computational support for the idea that tilt aftereffects arise from
lateral interactions between adapting feature detectors, as has long
been surmised.  They also suggest that indirect tilt aftereffects
could result from the conservation of synaptic resources.  The model
thus provides a unified computational explanation of self-organization
and both direct and indirect tilt aftereffects in the primary visual
cortex.

-------------------------------------------------------------------------
LATERALLY INTERCONNECTED SELF-ORGANIZING FEATURE MAP IN HANDWRITTEN 
DIGIT RECOGNITION

Yoonsuck Choe
Department of Computer Sciences, The University of Texas at Austin.

Masters Thesis and Technical Report AI95-236, August 1995 (65 pages).
ftp://ftp.cs.utexas.edu/pub/neural-nets/papers/choe.thesis.tar.Z 

An application of biologically motivated laterally interconnected 
synergetically self-organizing maps (LISSOM) to off-line recognition of 
handwritten digit is presented. The lateral connections of the LISSOM map 
learn the correlations between units through Hebbian learning. 
As a result, the excitatory connections focus the activity in local 
patches and lateral connections decorrelate redundant activity on the map.
This process forms internal representations for the input that are 
easier to recognize than the input bitmaps themselves or the activation 
patterns on a standard Self-Organizing Map (SOM). The recognition rate 
on a publically available subset of NIST special database 3 with LISSOM 
is 4.0% higher than that based on SOM, and 15.8% higher than that based 
on raw input bitmaps. These results form a promising starting point for 
building pattern recognition systems with a LISSOM map as a front end.