AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Differentiable JPEG-based Input Perturbation for Knowledge Distillation Amplification via Conditional Mutual Information Maximization

SIYU CHEN, Kaixiang Zheng, Ahmed H. Salamah, EN-HUI YANG

ICLR 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives

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
replace CE teacher with DJIP teacher (JPEG input perturbation)KD students · CIFAR-100, ViT-S -> DeiT-Ttop-1 accuracy: 69.86 -> 73.97Improved
replace CE teacher with DJIP teacherKD, WRN-40-2 -> WRN-16-2 · CIFAR-100top-1 accuracy: 74.92 -> 75.64Improved
replace CE teacher with DJIP teacherKD, ResNet-110 -> ResNet-20 · CIFAR-100top-1 accuracy: 70.67 -> 71.65Improved
replace CE teacher with DJIP teacherKD, ResNet-34 -> ResNet-18 · ImageNettop-1 accuracy: 70.66 -> 71.65Improved
replace CE teacher with DJIP teacherAT, ResNet-34 -> ResNet-18 · ImageNettop-1 accuracy: 70.70 -> 70.78No clear effect
larger teacher-student capacity gap (different architectures)DJIP teacher · CIFAR-100accuracy gain: up to 2.44% (different) vs smaller gains (same architecture)Improved
add the differentiable JPEG layer to MCMI (fixed centroids) (ablation)JMCMI · CIFAR-100 WRN-40-2 -> WRN-16-2 KDtop-1 accuracy: 74.92 -> 75.53Improved
add alternating centroid update on top of JMCMI (ablation)DJIP · CIFAR-100 WRN-40-2 -> WRN-16-2 KDtop-1 accuracy: 75.53 -> 75.64Improved
add DJIP JPEG layer to MCMI teacherDJIP-MCMI · CIFAR-100 VGG-13 -> VGG-8top-1 accuracy: 73.83 -> 74.26; gains of up to 0.92%Improved
increase lambda from 0.4 to 1.4 (ablation)DJIP teacher · ImageNet ResNet-34 -> ResNet-18 KDstudent accuracy: 71.362 -> 71.654Improved
increase lambda from 1.4 to 4.0 (ablation)DJIP teacher · ImageNet ResNet-34 -> ResNet-18 KDstudent accuracy: 71.654 -> 70.694Hurt
increase quantization rounds from 1 to 5 (ablation)DJIP teacher · ImageNet ResNet-34 -> ResNet-18 KDstudent accuracy: 71.278 -> 71.654Improved

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