Spiking Synchrony as a Learning Signal for Spiking Neural Networks

Abstract

Spiking neural networks promise high energy efficiency, yet training deep SNNs with timing-based local rules remains unstable and hard to scale. This thesis proposes SSDP, a population-level learning rule that strengthens synapses in proportion to the temporal co-firing of pre- and post-synaptic activity aggregated over a mini-batch. SSDP operates on group synchrony with a continuous Gaussian kernel and requires only binary spike indicators and first-spike indices. The implementation avoids per-synapse eligibility by computing first-spike times once per channel, making it compute- and memory-efficient and friendly to neuromorphic deployment. To improve credit assignment, we introduce DA-SSDP, which gates the synchrony update with a loss-negative feedback signal. In practice, SSDP and DA-SSDP act as auxiliary learners that run alongside backpropagation, injecting a biologically local learning signal toward synchronous firing while leaving the model's structure unchanged. Across static vision, event-based vision, and auditory temporal benchmarks, SSDP consistently improves accuracy and robustness without increasing model size. Analyses of spike timing jitter and feature show that SSDP sharpens class-relevant synchrony and enhances the model’s generalisation capacity. These results demonstrate that population-level synchrony is a practical and scalable learning mechanism for deep SNNs. The approach bridges biological plausibility and engineering efficiency, paving the way for future neuromorphic systems.