Optical neural network for unequally distributed neuron states: algorithm and application

Abstract

An adaptive clipping algorithm with asymmetric clipping points is proposed. By introducing the Gram-Schmidt orthogonal projection algorithm into the Hopfield model and then using the asymmetric adaptive clipping method, we have constructed an asymmetric adaptively clipped model. statistical simulations have revealed that the asymmetric adaptively clipped model has relatively strong performance for patterns with unequally distributed neuron states and of great similarity. since there are only three values used in the interconnection weights for the asymmetric adaptively clipped model, the interconnections are much more easily implemented with the optical technique than in the original Hopfield model. A 2-D hybrid optical neural network with 1024 neurons, in which a beam-direction encoding method is employed, is proposed to implement the adaptively clipped model. Preliminary experimental results for the reconstruction of traffic signs are shown.