A2SG: New training method makes spiking neural networks faster and more accurate

Training deep spiking neural networks (SNNs) has long been plagued by sharp loss landscapes and temporal inconsistencies caused by surrogate gradient methods. A new paper accepted at ICML 2026, authored by Yechan Kang and ten colleagues, tackles these issues head-on with A2SG (Adaptive and Asymmetric Surrogate Gradients). The adaptive component dynamically adjusts the effective window for spatio-temporal adaptation, reducing spatial gradient variation and maintaining directional consistency over time. The asymmetric component assigns larger gradients to neurons with higher membrane potentials, reflecting actual neuronal dynamics and yielding lower variation than symmetric surrogates.

The researchers prove a direct connection between local gradient variation and loss landscape curvature, showing that A2SG promotes convergence to flatter minima and improves generalization. Extensive experiments on CNN-based and Transformer-based SNNs across static image classification, neuromorphic datasets, and segmentation confirm that A2SG consistently improves both accuracy and energy efficiency. The code is publicly available, making this a general, reliable solution for the SNN community.