HIGH-CAPACITY NEURAL NETWORKS ON NONIDEAL HARDWARE

We present a new training-out algorithm for neural networks that permi ts good performance on nonideal hardware with limited analog neuron an d weight accuracy. Optical neural networks are emphasized with the err or sources including nonuniform beam illumination and nonlinear device characteristics. We compensate for processor nonidealities during gat ed learning (off-line training); thus our algorithm does not require r eal-time neural networks with adaptive weights. This permits use of hi gh-accuracy nonadaptive weights and reduced hardware complexity. The s pecific neural network we consider is the Ho-Kashyap associative proce ssor because it provides the largest storage capacity. Simulation resu lts and optical laboratory data are provided. The storage measure we u se is the ratio M/N of the number of vectors stored (M) to the dimensi onality of the vectors stored (N). We show a storage capacity of M/N = 1.5 on our optical laboratory system with excellent recall accuracy, > 95%. The theoretical maximum storage is M/N = 2 (as N approaches inf inity), and thus the storage and performance we demonstrate are impres sive considering the processor nonidealities we present. Our technique s can be applied to other neural network algorithms and other nonideal processing hardware.