Noise-Resilient Photonic Analog Neural Networks

The explosion of generative artificial intelligence (AI) has led to an unprecedented demand for AI accelerators. Photonic computing holds promise in this direction, offering speedups in bandwidth and latency. However, photonic integrated circuits (PICs) and their periphery input/output (I/O) components tend to be noisy due to the nature of analog computing. This can lead to accuracy degradation if not accounted for properly. In this paper, we characterize the typical noise levels present in photonic hardware accelerators for deep neural networks (DNNs). We explore several techniques including knowledge distillation, stability training, and standard Gaussian noise injection to improve the robustness of photonic DNNs. We validate our methods by training a Resnet model on the CIFAR-10 dataset and comparing the simulated test accuracy with different noise levels and image distortions. The robust training techniques discussed in this paper combined with the noise analysis of PICs provide a blueprint for robust photonic AI inference accelerators.