Exact Event-Driven Training for Efficient Neuromorphic Computing

This paper introduces an analytical, event-driven learning framework for training spiking neural networks (SNNs) that computes exact gradients for synaptic weights, transmission delays, and adaptive firing thresholds, improving accuracy and efficiency.

Details

Spiking neural networks (SNNs) offer potential efficiency gains over traditional neural networks by communicating with sparse, event-driven spikes rather than dense numerical activations. However, most SNN training pipelines rely on surrogate-gradient approximations or require dense time-step simulations, which conflict with the constraints of neuromorphic hardware and blur precise spike timing. This paper introduces an analytical, event-driven learning framework that computes exact gradients for synaptic weights, programmable transmission delays, and adaptive firing thresholds - three orthogonal temporal controls that jointly shape SNN accuracy and robustness. By propagating error signals only at spike events and integrating subthreshold dynamics in closed form, the method eliminates the need to store membrane-potential traces and reduces on-chip memory traffic by up to 24x. Across multiple sequential event-stream benchmarks, the framework improves accuracy by up to 7% over a strong surrogate-gradient baseline, while sharpening spike-timing precision and enhancing resilience to injected hardware noise.