Hybrid optical-electronic neural network with pseudoinverse learning for classification inference

Abstract

Constructing optical neural networks as analog artificial neural networks becomes a research highlight of implementing machine learning in a physical system. However, the training of an optical neural network with backpropagation error correction requires a complicated and less-efficient computational process. Here, we introduce a hybrid optical-electronic neural network to produce efficient artificial learning of handwritten digits without the backpropagation process. In the proposal, one can illuminate an input image with incident light and map the input image to a feature vector according to the transmitted light intensity. One can then encode the feature vector as an array of light and use an optical matrix multiplication unit to multiply the feature vector by a learning weight matrix. This learning weight matrix is realized by the spatial light modulator, which is constructed from the pseudoinverse learning method. The output light intensity through the optical multiplication unit represents the recognition result of the input image. The proposed neural network without backpropagation achieves sufficient accuracy of handwritten digits classification, exposing the advantages of training speed acceleration and energy efficiency improvement.