Abstract :
Integer-weight neural nets (IWNN) are better suited for hardware implementation than their real-weight analogues. The authors present a learning procedure for generating multilayer IWNNs having all weights in the set {-3, -2, -1, 0, 1, 2, 3}. The performance of this procedure was evaluated on XOR, encoder/decoder and the MONK benchmark. The IWNNS were found to be as capable as their real-weight counterparts with regard to generalisation performance
