Vertically hierarchical electro-photonic neural network by cascading element-wise multiplication

Abstract

Integrated photonic neural networks (PNNs) usually adopt traditional convolutional neural network (CNN) or multilayer perceptron (MLP) network models. These models consist of horizontally cascaded deep layer architectures interleaved by nonlinear activation functions. However, there are practical challenges for on-chip realizing such architectures, including the optical loss and the lack of efficient on-chip optical activation nonlinearity. Here, we propose a vertically hierarchical photonic neural network leveraging electro-optical element-wise multiplication to extract an element-wise feature in a polynomial projection space, which enables high-accuracy classification. For this network architecture, the light propagates through only two fully connected linear layers; thus, vertical extension to the deep layer is not limited by optical loss. This electro-photonic network can perform equivalently to or outperform optical CNN and MLP models even without interleaving deep layers by activation functions, benchmarking ∼97.9%, ∼87.7%, and ∼90.3% average blind-testing accuracies, for the whole test sets of MNIST handwritten digits, Fashion-MNIST images, and KMNIST Japanese cursive characters, respectively. It also demonstrates a >99% accuracy for boundary prediction of 12-labeled clusters. This work presents a different PNN architecture, which offers both high performance and better amenability to an integrated photonics platform.