Analysis of the Hardware Imprecisions for Scalable and Compact Photonic Tensorized Neural Networks

We simulated tensor-train decomposed neural networks realized by Mach-Zehnder interferometer-based scalable photonic neuromorphic devices. The simulation results demonstrate that under practical hardware imprecisions, the TT-decomposed neural networks can achieve >90% test accuracy with 33.6× fewer MZIs than conventional photonic neural network implementations.