Abstract

The Hamming net is a neural network implementation of the optimum image classifier.¹ It is essentially a two-layer neural network. The first layer calculates the Hamming distance between the input pattern and each exemplar, i.e., classes. The second layer, known as MAXNET, selects the node with the maximum output. We have developed a modified Hamming net model in which the effect of the Hamming distance is enlarged. The modified model alleviates the dynamic range requirement to the spatial light modulators and reduces the iteration cycles in the MAXNET. The optical Hamming net is constructed based on an architecture using a lenslet array and inexpensive liquid crystal televisions (LCTVs).² The input patterns as well as the interconnection weight matrices (IWMs) in both layers are bipolar signals in the Hamming net. To accomplish the bipolar multiplication optically, the IWMs and the input patterns are area modulated before being displayed on the LCTVs. The experimental results have shown that the optical Hamming net has a larger information capacity than other previously reported optical neural networks.²

© 1991 Optical Society of America