Hardware Implementation of Speech Recognition in Noise-Resilient Photonic Spiking Neural Network With Rate-Coding

Photonic neuromorphic computing is a combination of optics and brain-like computing, offering a promising alternative to traditional von Neumann architectures by enabling high-speed, low-latency, and energy-efficient computation. The spiking neural network (SNN), inspired by the brain's spatiotemporal encoding and processing mechanisms, employs sparse, event-driven spikes to achieve low power consumption. However, the inherent instability of optical devices presents challenges for precise computations, particularly in implementing both linear weight operations and nonlinear spiking activations. To address these challenges, this work explores algorithmic resilience as a pathway to enhance the robustness of photonic SNNs. We propose a noise-resilient photonic SNN architecture optimized for basic pattern classification tasks. Focusing on noise resilience and network performance, we adopt rate-coded SNNs trained using surrogate gradient methods, where noise is considered as a necessary factor in the algorithmic aspect. With rate coding, decisions are made based on the output spike rate or equivalently, the number of spikes, which is able to cope with nonlinear errors better compared to temporal coding. The proposed architecture is evaluated through online inference on the TIDIGITS dataset using a self-fabricated DFB-SA laser chip, and the final inference accuracy reaches 98.5%.