An in-the-loop training method for VLSI neural networks

This paper deals with the in-the-loop training of an intelligent sensor that is based on the use of an artificial neural network with analog neurons, programmable digital weights and an integrated photosensitive array. In the training method, each of the neuron activation functions of the actual physical realization is measured and then modeled in terms of a small neural network. These small neural networks are embedded in a larger neural network that models the complete neural network. For the training of the complete neural network model, an algorithm that allows one to train the network when the analytic nature of both the nonlinear neuron activation function and its derivative are not known is presented. We also describe an approach to training the hardware implementation using digital weights of low resolution where the weight quantization effects are especially evident