Abstract

We present an adaptive neural-network model in which each neuron unit has two inputs, a single potential, and two outputs. This network model can exhibit general functionality, although it uses only nonnegative values for all signals and weights in the interconnection. In this dual-function network (DFN), each output is a distinct nonlinear function of the potential of the neuron. Each neuron-unit input receives a linear combination of both outputs from any set of neurons in the preceding layer. A learning algorithm for multilayer feedforward DFN's based on gradient descent of an error function and similar to backward-error-propagation learning is presented.

© 1990 Optical Society of America