Micro-ring resonator assisted spiking neural network for efficient object detection.

Abstract

Optical computing and spiking neural networks (SNNs) have garnered significant attention as next-generation technologies due to their high parallelism and low-energy consumption. However, the current implementations for realizing spiking neurons of photonic neuromorphic computing mainly rely on active devices or nonlinear effects, which pose challenges for large-scale integration and energy conservation. Moreover, most existing optical SNN applications have been limited to simple image classification tasks. To address these limitations, we propose an optical-assisted SNN model based on the passive add-drop micro-ring resonator (ADMRR), which simulates the membrane potential accumulation in spiking neurons through optical temporal integration. System-level object detection is conducted numerically by the spiking version of the modified YOLO algorithm with ADMRR-based neurons. The results show that the proposed photonic SNN achieves performance exceeding 98% of that attained by computer-based SNN on the PASCAL VOC dataset, which contains 11,530 images across 20 object categories. Our work offers advantages including simplicity, enhanced parallelism, ease of large-scale integration, and effective emulation of neuronal leakage and integration dynamics, paving the way for the widespread use of photonic SNNs in more complex image processing tasks.