A training algorithm for discrete multilayer perceptrons

A learning algorithm for discrete multilayer perceptrons for binary patterns which guarantees convergence is introduced. Only two layers (one hidden layer) are required for binary patterns. Neurons in the hidden layer develop, as necessary, by learning without supervision. The computational amount is much less than that of the backpropagation algorithm. In the networks, neurons with hard limiters as their activation functions and integer weights and thresholds are used. Hence, accurate hardware implementation of trained networks can be easily realized using readily available VLSI technology.<>