Learning algorithms for optical multilayer neural networks

Two issues related to the optical implementation of feedforward multilayer neural networks are discussed. First, an anti-Hebbian local learning rule is introduced to eliminate the need for error backpropagation in network training, as required by the commonly used back error propagation algorithm. Second, the effect of using square-law detectors in optical neural networks is considered, and an orthogonal vector algorithm is proposed to train such networks. While the first algorithm significantly reduces the complexity of optical neural networks, the second algorithm enables use of simple square-law optical detectors to implement neurons. Preliminary simulations have shown that not only are these two algorithms capable of solving typical pattern recognition problems, but they are also computationally effective