Neural networks for high-storage content-addressable memory: VLSI circuit and learning algorithm

Abstract

Neural networks used as content-addressable memories show unequaled retrieval and speed capabilities in problems srreh as vision and pattern recognition. We propose a new implementation of a VLSI fully interconnected neural network with only two binary memory points per synapse. The small area of single synaptic cells allows implementation of neural networks with hundreds of neurons. Classical learning algorithms like the Hebb's rule show a poor storage capacity, especially in VLSI neural networks where the range of the synapse weights is limited by the number of memory points contained in each connectiorq we propose a new algorithm for programming a Hopfield neuraf network as a high-storage content-addressable memory. The storage capacity obtained with this algorithm is very promising for pattern recognition applications.

Analog implementations of artificial neural networks

“…al. which does modify the learning algorithm to accommodate the hardware [29]. The chip by Furman and Abidi is designed to be cascadable to produce multi-layer networks and the feedforward and BEP signals share the same physical inter-chip connections by using the voltages and currents as separate signals.…”

Design and fabrication of a GaAs weight-summation circuit WSC for use in neural networks

“…In conventional circuit design, a weight-summation circuit (WSC) will source or sink the currents generated in the multipliers directly to the input of an activation function circuit (Verleysen and Sirletti 1989) which also acts as the current-to-voltage converter. This type of design often suffers from excessive noise problems since the noise contribution from each of the multipliers will also be summed and will reduce the overall signal-to-noise ratio of the WSC.…”

Neural Networks in Analog Hardware — Design and Implementation Issues

“…A typical example of a mixed signal approach is the use of a local digital memory for storing the values of the weights. Such digital weight values can use just one 77,188 or several bits and a D/A converter. 46,90,150 Optical implementations of neural networks use optical encoding and processing of information.…”