AI Paper Atlas

Research map / Architectures and efficiency / Spiking networks

Zeroth-Order Forward-Only SNN Training Inspiring Neuromorphic On-Chip Learning

Mingyue Qin, Shuyu Yin, Qinghai Guo, Peilin Liu, Xiaolin Huang, Fei Wen

ICML 2026 · Architectures and efficiency › Spiking networks · approach: Spiking transformers

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply zeroth-order training to an SNN instead of an ANN (Heaviside spiking activation replacing ReLU)DeepZero (CGE) on ResNet20 · CIFAR10ZO training successHurt
train in a low-dimensional PCA subspace of the BP trajectory instead of full spaceBP on Spike-ResNet50 · ImageNet1Ktraining speed: fewer than 10 epochs of subspace training match 60 epochs of full-space BPImproved
use SZO-CGE instead of SZO-RGE in the low-dimensional subspace (ablation)SZO · training from scratchaccuracyNo clear effect
increase number of queries q in SZO-RGE (q=1 to q=20) (ablation)SZO-RGE · CIFAR10 Spike-ResNet20accuracy (%): 81.29 -> 83.71Improved
use SZO-CGE instead of full-space BP in continual trainingSpike-ResNet18 · CIFAR100 continual trainingmax accuracy (%): 74.42 -> 74.59Improved
use subspace-based SZO instead of Tent with full-space BP for test-time adaptation (Continual)Spike-ResNet50 · ImageNetC severity 5accuracy (%): 3.89 -> 33.37Improved
vary perturbation scale mu from 0.6 to 2.2 (ablation)SZO-CGE · CIFAR10 / CIFAR100 from scratchaccuracy (%): 85.91 -> 85.92 (CIFAR10, mu 0.6 -> 2.2)No clear effect
increase subspace dimension (ablation)SZO-CGE · CIFAR100 Spike-ResNet18 from scratchaccuracy (%): 70.97 -> 73.87 (dim 20 -> 60, first window)Improved
replace subspace ZO with feedback-based OPZO or DFA on deep ResNetsOPZO / DFA baselines · CIFAR100 Spike-ResNet18accuracy (%): 73.35 -> 36.90 (OPZO)Hurt
use SZO instead of BP for memorySpike-ResNet50 · ImageNet1K GPU memoryGPU memoryImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy