Theoretical studies of a dynamic neuro-vision processor with a biologically motivated design

A vision system for an autonomous intelligent machine must be quick, effortless and highly robust. At the moment only biological organisms have attained these visual capabilities. In sharp contrast, most machine vision systems are slow, consume excessive computational power and are intolerant to stimulus variability. In this dissertation a biological motivated design of a dynamic neural network called the scPositive- scNegative (scPN) neural processor is introduced. The computational transformations performed by the scPN neural processor for storing and processing spatio-temporal information are loosely based upon the functional dynamics exhibited by nervous tissue layers in the cortical and thalamic regions of the brain.

The scPN neural processor is a plausible hardware visual information processor that can be applied towards the development of sophisticated neuro-vision systems for engineering applications. The computational role of the scPN neural processor extends beyond the scope of most existing artificial neural networks because it can be programmed to generate multiple modes of dynamic behavior such as multiple stable states, transients and limit cycle oscillations. These distinct modes of dynamic behavior are useful for storing, processing and encoding visual information. In this context, the overall construction of a robust neuro-vision system involves a parallel and hierarchical architecture of numerous scPN neural processors that are individually programmed to perform specific visual tasks.

The steady-state behavior and temporal phenomena generated by the scPN neural processor are investigated by phase-plane analysis. A brief discussion of how this network can be applied to a neuro-vision design is presented. Extensive simulation studies are provided throughout to substantiate the concepts introduced in this work.