Abstract

Neural networks (NN's) for multiple-class pattern recognition (PR) offer an organized and efficient optimization procedure to produce nonlinear-discriminant decision surfaces. These NN techniques are automated and are needed for difficult PR problems. Many NN algorithms and applications have problems, such as requiring an excessive number of neurons, not using existing PR techniques, and requiring a number of ad hoc parameters to be selected (thus, not achieving the automated advantages of NN's). This paper offers algorithms and architecture solutions to these issues. The resultant NN combines NN/PR algorithms, plus analog/digital and optical/electronic technologies. 1 consider a three-layer NN (this can produce any piecewise linear decision surface) with analog input neurons and weights (optics provides this and is needed to realize its full advantages). The large number of interconnections possible optically is its major advantage. Initial weights (not random ones, as is conventional) and an automatically selected number of hidden-layer neurons are provided by means of PR techniques. The weights are then adapted by means of NN techniques. The number of input neurons is reduced by using feature-space and symbolic input neuron representations. A hybrid optical/digital architecture is described, and various applications of it (demonstrating its multifunctional nature) are noted.

© 1990 Optical Society of America