Fast Computation Using Multi-Zero Neural Networks

Abstract

The multi-zero artificial neural network was derived from a study of the stability and convergence properties of a feedback (or auto-associative) neural system. The nonlinear response function of neurons in the system is an odd polynomial (or a topologically similar) function of 2M+1 zeros with odd zeros equal to a set of consecutive integers. If the connection matrix is programmed correctly, the system will then perform stable operations exhibiting the following characteristics. 1. The system will transform any N-bit analog input to an N-bit, M-ary (or M-valued), digital output. 2. The output will be locked-in when the input is removed. It will be changed to another locked-in digital vector when it receives another input. 3. The speed is fast because the circuit is free-running, parallel, and M-ary. The accuracy is high because the computation is digital. Because of these unique properties, the network can be used in the design of a fast computing system. This paper reports the origin of this multi-zero system, the analysis of its properties, and the design of a fast, M-ary, digital multiplier using this system.