AI Paper Atlas

Research map / Optimization / Optimizers

SURGE: Surrogate Gradient Adaptation in Binary Neural Networks

Haoyu Huang, Boyu Liu, Linlin Yang, Yanjing Li, Yuguang Yang, Xuhui Liu, Canyu Chen, Zhongqian Fu, Baochang Zhang

ICML 2026 · Optimization › Optimizers · approach: Constrained and non-backprop training · 0 citations (OpenAlex)

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
add Dual-Path Gradient Compensator (ablation)binarized ResNet-20 baseline · CIFAR-10accuracy: 87.4 -> 87.8Improved
add Adaptive Gradient Scaler on top of DPGC (ablation)SURGE ResNet-20 · CIFAR-10accuracy: 87.8 -> 88.0Improved
use fixed scaling factor instead of adaptive scaling (ablation)SURGE ResNet-20 · CIFAR-10accuracy: fixed 0.07: 70.32% vs adaptive 87.95%Hurt
compensate only gradients outside the clipping range (ablation)SURGE ResNet-20 · CIFAR-10accuracy: 87.4 -> 87.7Improved
compensate only within-range gradients (ablation)SURGE ResNet-20 · CIFAR-10accuracy: 87.4 -> 87.6Improved
compensate all gradients (clipped and unclipped) (ablation)SURGE ResNet-20 · CIFAR-10accuracy: 87.4 -> 88.0Improved
add explicit alignment loss between auxiliary and binary branch outputs (ablation)BiT+SURGE · GLUE devaverage score: 72.0 -> 71.2Hurt
add SURGE to BiT under same wall-clock budgetBiT · GLUE dev (fixed wall-clock budget)average score: 65.2 -> 66.7Improved
add SURGE to BiTBiT BERT-base · GLUE devaverage score: 70.6 -> 72.0Improved
add SURGE vs best prior on one-stage ImageNetbinarized ResNet-18 · ImageNet-1K (one-stage training)top-1 accuracy: 61.4 -> 62.0Improved
add SURGE vs best prior on two-stage ImageNetbinarized ResNet-18 · ImageNet-1K (two-stage training)top-1 accuracy: 66.5 -> 66.7Improved
use 1x1 instead of 3x3 auxiliary convolutions (SURGE*) (ablation)SURGE ResNet-18 · ImageNet-1K (10 epochs)accuracy: 47.28 -> 47.21No clear effect

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