Modeling Self-Organization in the Visual Cortex

Abstract

This chapter talks about how the self-organizing map (SOM) architecture was originally motivated by the topological maps and the self-organization of sensory pathways in the brain. The standard SOM algorithm, where a global supervisor finds the maximally responding unit and determines the weight-change neighborhood, is a computationally powerful abstraction of the actual biological mechanisms. Such an abstraction has turned out extremely useful in many real-world tasks from visualization and data mining to robot control and optimization. To gain insight into biological knowledge organization and development, the self-organizing map architecture is extended with anatomical receptive fields, lateral connections, Hebbian adaptation, and spiking neurons. The resulting RF-SLISSOM model shows how the observed receptive fields, columnar organization, and lateral connectivity in the visual cortex can arise through input-driven self-organization, and how such plasticity can account for partial recovery following retinal and cortical lesions. The self-organized network forms a redundancy-reduced sparse coding of the input that allows it to process massive amounts of information efficiently, and explains how various low-level visual phenomena such as tilt aftereffects, and segmentation and binding can emerge. Such models allow understanding biological processes at a very detailed computational level, and are likely to play a major role in cognitive neuroscience in the future.