Optical processing and hybrid neural nets

摘要

Four different optical processing and hybrid neural net (NN) algorithms and architectures are described that show different uses for optical processing in NNs for multiclass distortion- invariant pattern recognition. The hybrid associative processor (AP) shows the use of optical feature spaces, feature space neurons, the use for linear algebra algorithms in NN design, and optical matrix-vector processors. For more difficult problems requiring nonlinear piecewise decision surfaces, the AP is replaced by a multilayer NN. The NN used combines pattern recognition and NN techniques and uses a digital/optical NN for hybrid training/classification. A symbolic correlator production system NN is described when multiple objects are present in the field of view. This uses an optical correlator, neurons that represent facts, and a production system NN that can be implemented optically or digitally. A new optimization NN formulation allows an all optical (or hybrid optical/electronic) implementation with a fixed interconnection mask and optical feedback.