Design, analysis, and gallium arsenide based implementation of a class of neural networks with application to vision and optoelectronics

TLDR: The implemented network's ability to perform contrast enhancement, data compression, and adaptation to mean input strength level, as well as its tunability of sensitivity, are well-suited for processing of sensory information in general and visual information in particular.

Abstract: A framework for extremely compact, all analog, and massively parallel implementation of shunting recurrent and nonrecurrent neural networks that is applicable to a wide variety of FET-based integration technologies is proposed. A specific circuit designed for implementation with gallium arsenide MESFETs is implemented. The implemented network's ability to perform contrast enhancement, data compression, and adaptation to mean input strength level, as well as its tunability of sensitivity, are well-suited for processing of sensory information in general and visual information in particular. Thus, if integrated with an array of photo-detectors, it can find widespread optoelectronics applications. These properties are also desirable for feature extraction for content addressable memory (CAM) systems. Furthermore, the network fits in self-organizing multilayer network architectures of Adaptive Resonance Theory. The asymptotic global stability of the network is shown by finding a global Liapunov energy function, which serves to show that the network itself is capable of CAM. The time trace of the local activity of the implemented network is found via a variational analysis which leads to the identification of higher order kernels in a Volterra series expansion. The same mathematical techniques are applied to find higher order spatial kernels which explains the extraordinary high order classification properties embedded in simple multiplicative nonlinearity.