Visual Perceptual Grouping in a Self-Organizing Map of Spiking Neurons

Yoonsuck Choe

Building artificial vision systems that understand the visual environment is a
very difficult task. Understanding the computational principles employed by
biological vision systems can give us powerful insights into building better
vision systems. Perceptual grouping is well-suited for such an investigation
because it bridges the levels of cognitive behavior and neural implementation.
In perceptual grouping tasks, features in the scene must be grouped (bound) into
objects and these objects must be segmented to form distinct representations.
The main questions are, (1) what are the mechanisms underlying grouping, and
and (2) how do they come to exist during development?

Based on experimental evidence, two separate theories have been proposed to
address these issues: (1) unsupervised, input-driven self-organization and (2)
temporal coding based on synchronized and desynchronized neural activity.  In
my work, these two components are brought together into a unified computational
model of the visual cortex, where the structure of the neural network is
learned through self-organization, and the temporal firing patterns are formed
to represent perceptual events. The self-organized afferent and lateral
connections in the model closely approximate neurophysiological data.  In
simulated contour integration experiments, the model behavior closely
replicates human performance, and demonstrates that such performance could be
due to self-organized lateral connectivity.