A Multi-Objective Particle Swarm Optimization Pruning on Photonic Neural Networks

Abstract

Motivated by the increasing capability of artificial intelligence (AI) in solving a large class of problems, integrated photonic neural networks (PNNs) with Mach-Zehnder Interferometers (MZIs), have shown some advantages, such as low power, low latency, and high bandwidth alternatives to digitally electric neural networks. However, as the complexity of the problem being tackled grows, PNNs are accompanied by massive model sizes, necessitating significant computational and tuning power consumption. To enable the deployment of large-scale PNNs in power-constrained environments and maintain inference performance, in this paper, we target at pruning the redundant phase weights in PNNs. More specifically, we first investigate the feasibility of pruning based on phase size and point out sparse pruning schemes. Additionally, a multi-objective PNNs pruning method that trade-offs the accuracy and the tuning power consumption of networks is proposed and we solve this model with swarm optimization, named PP-MOPSO. Experimental results demonstrate that PP-MOPSO achieves a 97.3% reduction in on-chip power consumption while maintaining 90.51% inference accuracy on PNNs using the MNIST dataset. In the other case with CIFAR-10 dataset, the method achieves 94.98% power savings with a 6.39% accuracy loss.