Multilayer optical-electrical spiking neural network with sparse spike event for speech recognition based on a...

Spiking neural networks (SNNs) are gaining increasing attention in artificial intelligence and machine learning due to their high power efficiency, which stems from their event-driven nature and sparse spiking activity. Optical platforms offer significant speed improvements for neural network computations. However, they face challenges in achieving low-threshold nonlinear activation, a crucial requirement for high expressivity in neural networks. Achieving nonlinear activation within the optical domain is essential to minimize redundant optical-electrical conversions. In this work, we propose the implementation of a multi-layer SNN with extremely sparse spike events for speech recognition on an optical-electrical platform. Notably, each neuron in SNN emits an average of 0.4 spikes with the time-to-first spike encoding strategy, which is about four times fewer spikes than in other spiking neural networks. Moreover, a self-fabricated distributed feedback laser with a saturable absorber (DFB-SA) laser is employed to perform nonlinear activation, yielding promising accuracy results with an accuracy of 90.5%. This study demonstrates the feasibility of nonlinear computation within the optical domain for processing a benchmark speech recognition dataset with sparse spike events, and paves the way for the application of photonic SNNs in more complex tasks.