Quantifying knowledge during full-layer ANN-to-SNN knowledge distillation

Di Hong

doi:10.1016/j.patcog.2026.113066

Quantifying knowledge during full-layer ANN-to-SNN knowledge distillation

2026-01-12

0

PRE

AI

DOI：10.1016/j.patcog.2026.113066

Original

Original request for help

Save

Abstract

En 中文

• Memory-Efficient Full-Layer Distillation: We introduce a similarity matrix to distill knowledge from all intermediate ANN layers to SNNs efficiently, significantly reducing memory and computation costs from (b × c × w × h)2 to b × b, thus enabling scalable and parallel layer-wise learning. • Quantitative Interpretability: Three novel metrics provide pixel-level insights into improved discriminative ability, learning dynamics, and optimization paths. • State-of-the-Art Performance: Achieves top-1 accuracy of 95.84% on CIFAR-10, 78.72% on CIFAR-100, and 68.37% on ImageNet, surpassing previous SNN methods.

Quantifying knowledge during full-layer ANN-to-SNN knowledge distillation

Abstract

Journal

Organization