Artificial neural networks using MOS analog multipliers

A neural network implementation that uses MOSFET analog multipliers to construct weighted sums is described. The scheme permits asynchronous analog operation of Hopfield-style networks with fully programmable digital weights. This approach avoids the use of components that waste chip area of require...

Full description

Saved in:
Bibliographic Details
Published in:IEEE journal of solid-state circuits 1990-06, Vol.25 (3), p.849-855
Main Authors: Hollis, P.W., Paulos, J.J.
Format: Article
Language:English
Subjects:
Analytical models
Artificial neural networks
MOSFET circuits
Neural networks
Neurons
Resistors
Symmetric matrices
Testing
Very large scale integration
Voltage
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:A neural network implementation that uses MOSFET analog multipliers to construct weighted sums is described. The scheme permits asynchronous analog operation of Hopfield-style networks with fully programmable digital weights. This approach avoids the use of components that waste chip area of require special processing. Two small chips have been fabricated and tested-one implementing a fully connected (recursive) network and the other containing isolated portions of a neuron. The fully connected network chip successfully solves simple graph partitioning problems, in confirmation of network simulations performed using an analytic model of the analog neuron. This result verifies the operation of the complete network, including common-mode biasing circuits and connection weight data paths. A direct scaling of this chip would allow the complete integration of 81-neuron fully connected networks with 6-b plus sign connection weights using 1.25- mu m design rules on a 1-cm die.< >
ISSN:0018-9200
1558-173X
DOI:10.1109/4.102684