!Converted with LaTeX2HTML 95.1 (Fri Jan 20 1995) by Nikos Drakos (email@example.com), CBLU, University of Leeds >
For over two decades, it has been known that the response properties of neurons in the primary visual cortex (V1), such as selectivity to oriented edges, are shaped by visual experience [4,5,22,23,32]. The discovery by von der Malsburg (; see also  and ) that simple computational rules could drive the development of oriented receptive fields (RFs) from visual input, raised the hope that much of the structure and function of V1 could be understood in terms of very simple neuronal behavior. However, since then substantial new discoveries have changed our understanding of the primary visual cortex. New evidence indicates that cells in V1 are coupled by a highly-specific structure of long-range lateral connections [15,16]. The patterns of lateral connections are not uniform or genetically determined, but develop based on visual experience [8,9,14,25,29] The connections are initially widespread, but develop into clustered patches at approximately the same time as the orientation maps form. In the adult, they primarily connect areas with the same orientation preference. It is not known how these lateral connections develop, why their patterns are related to the response properties of cortical cells, and what role they play in visual processing.
New discoveries have also changed the notion that receptive fields in the adult cortex behave as static filters. The RFs are highly mutable, and can expand in size dramatically when occluded for a short time . The effects of such expansion are visible in psychophysical experiments at time scales as short as one second . Therefore, the primary visual cortex appears to be a continously adapting system in a dynamic equilibrium with the input environment.
Computational models can play a fundamental role in understanding the development and function of such dynamic systems. Cortical function depends upon the collective behavior and interactions of large numbers of neurons. With the advent of massively parallel computers in the last five years, it has become possible to simulate such large neural systems. At the same time, neurobiological techniques to probe into the brain have reached the level of sophistication necessary to constrain and validate such models. This technological confluence provides a timely opportunity to explore the fundamental principles of cortical self-organization through large-scale computational experiments.
Several earlier computational models have shown how receptive fields (such as selectivity to different orientations) and their global organization in the cortical network can develop through Hebbian self-organizing mechanisms [10,17,27,30,31,34,44]. However, so far such models have assumed that lateral interactions between cells have a fixed, simple form (such as a Gaussian, or bell-shaped, profile) and have focused on explaining how cortical and retinal activity organize the afferent synaptic weights. To date, the self-organization of lateral connections in the visual cortex and their role in visual processing has not been investigated computationally. Furthermore, none of the developmental models have attempted to explain the dynamic nature of mature receptive fields. To provide a cogent theory, a computational model of cortical development should demonstrate how both afferent and lateral connections can organize cooperatively and simultaneously, and how the resulting structure can be dynamically maintained in a continuously-adapting equilibrium with the visual input.
This article shows that Hebbian self-organization in a large recurrent network of simple neural elements can provide such a unified account. The model is based on two fundamental assumptions: (1) The lateral connections are primarily excitatory in the short range and inhibitory in the long range, and (2) The afferent and lateral synapses adapt by the same Hebbian mechanism. Based on these assumptions, the model explains computationally how (1) the receptive fields develop orientation preference and selectivity, (2) such receptive fields are organized into columns and structures such as linear zones, pinwheel centers and fractures, (3) lateral connections self-organize and follow the global organization of receptive fields, and (4) the resulting receptive field structures maintain a dynamic equilibrium with the input. The model also suggests a functional role for the lateral connections: during development, they learn the activity correlations between cortical neurons, and during visual processing, filter out these correlations from cortical activity to form a redundancy-reduced sparse coding of the visual input.