Abstract

Objects are usually stored in neural networks as sets of weights on synapses that multiply input values. For convenience, the stored and input values are often set equal to the binary values 1 and 0. Using instead 1 and –1, as in the Hamming model, often yields improved network performance. However, because input objects do not usually contain negative values, input values of 1 and –1 ordinarily require preprocessing of the input data.

© 1990 Optical Society of America