New Method Converts Direct-Coded SNNs to Energy-Efficient Event-Based Models

Spiking Neural Networks (SNNs) promise ultra-low-power computation on neuromorphic hardware, but a critical divide has emerged between training paradigms. Direct coding — which enables backpropagation via continuous-valued surrogate activations — yields strong performance but consumes significantly more energy than purely event-based SNNs. This energy gap has limited the practical deployment of direct-coded models, even as large pretrained SNN databases accumulate. A team led by Nhan Trong Luu from Vietnam's Duy Tan University and other institutions now tackles this underexplored bottleneck in a new paper (arXiv:2605.07207, accepted at IEEE Signal Processing Letters 2026).

Their work presents the first systematic analysis of the direct-to-event transfer problem. They identify fundamental challenges — such as mismatched spiking thresholds, loss of temporal dynamics, and activation distribution shifts — that degrade performance when naively converting a direct-coded SNN to event-based operation. The researchers then propose a suite of methods: threshold calibration, surrogate-to-event mapping layers, and fine-tuning strategies that maintain task accuracy while enabling energy-efficient spike-driven computation. The approach makes it feasible to leverage existing direct-coded SNN repositories (trained on vision, audio, or sensor data) on neuromorphic chips without retraining from scratch. This could accelerate deployment in edge AI, robotics, and always-on IoT devices where power budgets are tight. The authors demonstrate promising results on standard benchmarks, though full energy measurements and hardware tests remain for future work.