17.1 Neural Networks and Analog VLSI

Artificial neural networks, or neurocomputers, provide an alternative form of computation that attempts to mimic the functionality of the human brain[1]. These networks seem to be better suited for information processing applications and tasks, such as optimization, pattern recognition and associative recall, than traditional digital computers.

Artificial neural networks have experienced significant growth in the last few years. However, only very large scale integration (VLSI) can realize the true computing potential of massively parallel neural networks. The realization of these neurocomputers, which are optimized to the computation of neural models, follows one of two approaches[2]:

1. general-purpose neurocomputers that consist of programmable processor arrays for emulating a range of neural network models, or,

2. special-purpose neurocomputers that are dedicated hardware implementations of a specific neural network model. However, any programmable neurocomputer is an order of magnitude slower than what could be achieved by directly fabricating a neural network in hardware. For this reason, far more dedicated neural hardware is being developed than programmable neural processors.

Technologies used in special purpose neural network implementation are broadly categorized into silicon ^{[3 5]}, using analog or digital or mixed analog/digital integrated circuits, and optical or electro-optical ^{[6 7]}.

Neural network designers could theoretically choose any neural model for implementation, although researchers favor the Kohonen[8] or the Hopfield[9] associative memory models because of their extreme simplicity. In fact, these simple models lend themselves naturally to analog electronic implementation. Furthermore, some of the traditional analog design...