Weight and neuron output resolution requirements in optoelectronic neural networks

Abstract

We demonstrate that the resolution requirements of the optoelectronic devices used in the communication links of an analog multiperceptron neural network, trained with the standard backpropagation algorithm, can be simultaneously reduced to 8 bits (receiver) and 4 bits (transmitter), respectively, without any significant effect on the network's learning or generalization performances. In addition, we also show that a simple modification to the sigmoidal function, used within each neuron architecture, permits the resolution requirements of the optoelectronic receiver to be further reduced to 4 bits without any additional effect on network performance other than a reduction in learning rate. Both of these limited device resolution performances, however, can be achieved only provided that the weight-storage and the weight-updating procedures are maintained at 14 bits or greater.