Image Classification on CIFAR-10 (test) (Acc and AA)
96.03AccuracySADAJEM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| SADAJEMcategory=JEM variants2025.05 | 96.03 | 29.63 | — | |
| EL2NPruning ratio=10%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.69 | — | — | |
| IF-BetaPruning ratio=30%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.69 | — | — | |
| ForgettingPruning ratio=10%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.66 | — | — | |
| IF-BetaPruning ratio=10%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.66 | — | — | |
| DUALPruning ratio=10%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.58 | — | — | |
| ForgettingPruning ratio=30%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.54 | — | — | |
| DUALPruning ratio=30%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.54 | — | — | |
| EL2NPruning ratio=30%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.53 | — | — | |
| ForgettingPruning ratio=50%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.51 | — | — | |
| IF-BWSPruning ratio=30%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.51 | — | — | |
| IF-BWSPruning ratio=10%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.47 | — | — | |
| RandomPruning ratio=10%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.44 | — | — | |
| IF-CCSPruning ratio=10%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.4 | — | — | |
| GraNdPruning ratio=10%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.39 | — | — | |
| EL2NPruning ratio=50%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.38 | — | — | |
| IF-BetaPruning ratio=50%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.38 | — | — | |
| DUALPruning ratio=50%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.32 | — | — | |
| MediumPruning ratio=10%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 95.12 | — | — | |
| Medium*Pruning ratio=10%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.97 | — | — | |
| RandomPruning ratio=30%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.89 | — | — | |
| IF-BWSPruning ratio=50%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.82 | — | — | |
| IF-CCSPruning ratio=30%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.65 | — | — | |
| GraNdPruning ratio=30%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.45 | — | — | |
| RandomPruning ratio=50%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.15 | — | — | |
| MediumPruning ratio=30%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.14 | — | — | |
| Medium*Pruning ratio=30%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 94.02 | — | — | |
| IF-BetaPruning ratio=70%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.9 | — | — | |
| ForgettingPruning ratio=70%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.86 | — | — | |
| DUALPruning ratio=70%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.85 | — | — | |
| IF-CCSPruning ratio=50%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.77 | — | — | |
| JEM++category=JEM variants2025.05 | 93.73 | 41.06 | — | |
| PoSFormerTraining=from scratch, Orthogonality loss coefficient (λo)=0.12026.05 | 93.4 | — | — | |
| PoSFormerTraining=from scratch, Orthogonality loss coefficient (λo)=1.02026.05 | 93.4 | — | — | |
| IF-CCSPruning ratio=70%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.39 | — | — | |
| EL2NPruning ratio=70%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.27 | — | — | |
| MediumPruning ratio=50%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 93.15 | — | — | |
| Medium*Pruning ratio=50%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 92.99 | — | — | |
| JEMcategory=JEM variants2025.05 | 92.9 | 4.28 | — | |
| BaselineTraining=from scratch2026.05 | 92.5 | — | — | |
| RandomPruning ratio=70%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 92.28 | — | — | |
| Ours (γ=0)Arch. (approx.)=MoE-Transformer, L=4, D=256, Params=~6–8 M, Strict FF=yes, Aug.=RandAug+RRC+CJ+Blur, TTA=yes, Seeds=32026.05 | 91.32 | — | — | |
| IF-BetaPruning ratio=80%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 91.3 | — | — | |
| GraNdPruning ratio=50%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 91.25 | — | — | |
| MediumPruning ratio=70%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 91.12 | — | — | |
| Medium*Pruning ratio=70%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 91.05 | — | — | |
| DUALPruning ratio=80%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 90.98 | — | — | |
| IF-CCSPruning ratio=80%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 90.87 | — | — | |
| Gong et al. (2025)Arch. (approx.)=VGG8 CNN, Params=n/r, Strict FF=yes, Aug.=RandCrop+Flip, TTA=no, Seeds=n/r2026.05 | 90.62 | — | — | |
| Ours (γ=0)Arch. (approx.)=MoE-Transformer, L=4, D=256, Params=~6–8 M, Strict FF=yes, Aug.=RandAug+RRC+CJ+Blur, TTA=no, Seeds=32026.05 | 90.5 | — | — | |
| EB-JDAT-SADAJEMBase Model=SADAJEM2025.05 | 90.39 | 66.3 | — | |
| EB-JDAT-JEM++Base Model=JEM++2025.05 | 90.37 | 64.61 | — | |
| IF-BWSPruning ratio=70%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 89.8 | — | — | |
| RandomPruning ratio=80%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 89.4 | — | — | |
| IF-BetaPruning ratio=90%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 88.51 | — | — | |
| Wu et al. (2024)Arch. (approx.)=CNN, L=10, Params=n/r, Strict FF=no (OG/DF), Aug.=RandCrop+Flip, TTA=no, Seeds=n/r2026.05 | 88.2 | — | — | |
| DUALPruning ratio=90%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 88.19 | — | — | |
| IF-CCSPruning ratio=90%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 88.18 | — | — | |
| Medium*Pruning ratio=80%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 87.92 | — | — | |
| CleanJPEG quality=902026.06 | 87.9 | — | — | |
| MediumPruning ratio=80%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 87.69 | — | — | |
| CleanJPEG quality=802026.06 | 87.67 | — | — | |
| CleanJPEG quality=702026.06 | 87.25 | — | — | |
| CleanJPEG quality=602026.06 | 86.6 | — | — | |
| IF-BWSPruning ratio=80%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 86.5 | — | — | |
| Sun et al. (2025)Arch. (approx.)=ResNet-like CNN, L=17, Params=~5–10 M, Strict FF=yes, Aug.=RandCrop+Flip, TTA=no, Seeds=n/r2026.05 | 86.22 | — | — | |
| IF-BWSPruning ratio=90%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 86.2 | — | — | |
| CleanJPEG quality=502026.06 | 86.09 | — | — | |
| CleanJPEG quality=402026.06 | 85.84 | — | — | |
| MediumPruning ratio=90%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 85.64 | — | — | |
| Aghagolzadeh and Ezoji (2025)Arch. (approx.)=ViT, D=240, L=7, Params=n/r, Strict FF=yes, Aug.=RC+RandAug, TTA=no, Seeds=n/r2026.05 | 85.4 | — | — | |
| Medium*Pruning ratio=90%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 85.39 | — | — | |
| JEATcategory=Energy-based AT2025.05 | 85.16 | 28.43 | — | |
| CleanJPEG quality=302026.06 | 85.05 | — | — | |
| GraNdPruning ratio=70%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 84.48 | — | — | |
| CleanJPEG quality=202026.06 | 84.4 | — | — | |
| RandomPruning ratio=90%, Retraining protocol=w/o retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 84.26 | — | — | |
| Dooms et al. (2024)Arch. (approx.)=VGG-like CNN, L=12, Params=~8.5 M, Strict FF=no (OLU), Aug.=RandCrop+Rot, TTA=no, Seeds=n/r2026.05 | 83.51 | — | — | |
| WEATcategory=Energy-based AT, initialization=PCA2025.05 | 83.36 | 49.02 | — | |
| WEATcategory=Energy-based AT, initialization=IF2025.05 | 83.36 | 49.02 | — | |
| LSPJPEG quality=202026.06 | 83.14 | — | — | |
| Sarode et al. (2026)Arch. (approx.)=14-layer CNN, Params=n/r, Strict FF=yes, Aug.=none, TTA=no, Seeds=32026.05 | 83.08 | — | — | |
| CleanJPEG quality=102026.06 | 82.24 | — | — | |
| LSPJPEG quality=102026.06 | 82.03 | — | — | |
| EMNJPEG quality=102026.06 | 81.93 | — | — | |
| LSPJPEG quality=302026.06 | 81.59 | — | — | |
| Chen et al. (2025)Arch. (approx.)=conv. + direct feedback, Params=n/r, Strict FF=no, Aug.=RandCrop+Flip, TTA=no, Seeds=n/r2026.05 | 80.8 | — | — | |
| EMNJPEG quality=202026.06 | 80.3 | — | — | |
| GraNdPruning ratio=80%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 80.01 | — | — | |
| LSPJPEG quality=402026.06 | 78.89 | — | — | |
| GUEJPEG quality=102026.06 | 78.88 | — | — | |
| BPLoss Function=CE2026.05 | 78.5 | — | — | |
| GUEJPEG quality=202026.06 | 78.24 | — | — | |
| EL2NPruning ratio=80%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 77.43 | — | — | |
| LSPJPEG quality=502026.06 | 76.57 | — | — | |
| EMNJPEG quality=302026.06 | 76.37 | — | — | |
| GraNdPruning ratio=90%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 75.63 | — | — | |
| BPLoss Function=MSE2026.05 | 75.2 | — | — | |
| GUEJPEG quality=302026.06 | 74.53 | — | — | |
| ForgettingPruning ratio=80%, Retraining protocol=w/ partial or full retraining, Teacher architecture=ResNet-18, Student architecture=ResNet-18, KD setting=true2026.06 | 74.4 | — | — |