A Comparison of Hamming and Hopfield Neural Nets for Pattern Classification

The Hopfield model neural net has attracted much recent attention. One use of the Hopfield net is as a highly parallel content-addressable memory, where retrieval is possible although the input is corrupted by noise. For binary input patterns, an alternate approach is to compute Hamming distances between the input pattern and each of the stored patterns and retrieve that stored pattern with minimum Hamming distance. We first show that this is an optimum processor when the noise is statistically independent from bit to bit. We then present a Hamming Neural Net which is a highly parallel implementation of this algorithm that uses computational elements similar to those used in a Hopfield net. We also compare the Hopfield and Hamming nets for several applications. For the cases considered, the Hamming net generally outperforms the Hopfield net. Also, the Hamming net requires fewer interconnects than the fully connected Hopfield net.