Image Classification on CIFAR-10 (test) (Accuracy and Sparsity)
98.2AccuracyDP-Adam
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DP-Adamepsilon=8, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 98.2 | — | — | — | — | — | |
| DP-GRAPEepsilon=8, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 98.2 | — | — | — | — | — | |
| DP-Adamepsilon=4, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 98.1 | — | — | — | — | — | |
| DP-GRAPEepsilon=2, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97.9 | — | — | — | — | — | |
| DP-GRAPEepsilon=4, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97.8 | — | — | — | — | — | |
| Naïve DP-GaLoreepsilon=8, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97.7 | — | — | — | — | — | |
| DP-Adamepsilon=2, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97.5 | — | — | — | — | — | |
| Naïve DP-GaLoreepsilon=4, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97.5 | — | — | — | — | — | |
| NASNet-ASearch Cost=22,400 GPU-days, Generation=Cell-based, Training Protocol=Traditional NAS (fully trained)2026.05 | 97.35 | — | — | — | — | — | |
| DARTSSearch Cost=1 GPU-day, Generation=Differentiable, Training Protocol=Traditional NAS (fully trained)2026.05 | 97.24 | — | — | — | — | — | |
| DP-Adamepsilon=1, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97.2 | — | — | — | — | — | |
| RW-SAMModel=WideResNet2025.09 | 97.11 | — | — | — | — | — | |
| ENASSearch Cost=0.5 GPU-days, Generation=RL + sharing, Training Protocol=Traditional NAS (fully trained)2026.05 | 97.11 | — | — | — | — | — | |
| EfficientNetGeneration=Scaling, Training Protocol=Traditional NAS (fully trained)2026.05 | 97.1 | — | — | — | — | — | |
| Naïve DP-GaLoreepsilon=2, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97 | — | — | — | — | — | |
| DP-GRAPEepsilon=1, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 97 | — | — | — | — | — | |
| SAMModel=WideResNet2025.09 | 96.91 | — | — | — | — | — | |
| LSFormerSOPs (G)=18.05, Power (mJ)=1.392026.05 | 96.73 | — | — | — | — | — | |
| SGDModel=WideResNet2025.09 | 96.47 | — | — | — | — | — | |
| ConfidenceAL Setup (Initial + Rounds x Batch Size)=100 + 10 x 202025.10 | 96.46 | — | — | — | — | — | |
| Naïve DP-GaLoreepsilon=1, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 96.4 | — | — | — | — | — | |
| HRPNoise mode=Sym., Noise ratio=20%2026.05 | 96.4 | — | — | — | — | — | |
| RW-SAMModel=ResNet-502025.09 | 96.34 | — | — | — | — | — | |
| QKFormerParam. (M)=6.74, T=42026.05 | 96.31 | — | — | — | — | — | |
| UncertParam. (M)=6.74, T=42026.05 | 96.3 | — | — | — | — | — | |
| RW-SAMModel=ResNet-182025.09 | 96.24 | — | — | — | — | — | |
| SWformerParam. (M)=7.51, T=42026.05 | 96.1 | — | — | — | — | — | |
| ICBMsBackbone=WideResNet2026.05 | 96.1 | — | — | — | — | — | |
| SAMModel=ResNet-502025.09 | 96.06 | — | — | — | — | — | |
| STAttenT=42026.05 | 96.03 | — | — | — | — | — | |
| PSSCLNoise mode=Sym., Noise ratio=20%2026.05 | 96 | — | — | — | — | — | |
| HRPNoise mode=Sym., Noise ratio=50%2026.05 | 96 | — | — | — | — | — | |
| SAMModel=ResNet-182025.09 | 95.99 | — | — | — | — | — | |
| SpikingformerSOPs (G)=14.81, Power (mJ)=1.142026.05 | 95.81 | — | — | — | — | — | |
| ERM + DReSBackbone=PARN18, Batch Size=128, Optimizer=SGD with momentum 0.92025.09 | 95.8 | — | — | — | — | — | |
| RRLNoise mode=Sym., Noise ratio=20%2026.05 | 95.8 | — | — | — | — | — | |
| HRPNoise mode=Asym., Noise ratio=40%2026.05 | 95.7 | — | — | — | — | — | |
| SGDModel=ResNet-502025.09 | 95.64 | — | — | — | — | — | |
| SGDModel=ResNet-182025.09 | 95.62 | — | — | — | — | — | |
| MixupBackbone=PARN18, Batch Size=128, Optimizer=SGD with momentum 0.92025.09 | 95.6 | — | — | — | — | — | |
| S-TransformerParam. (M)=10.28, T=42026.05 | 95.6 | — | — | — | — | — | |
| UNICONNoise mode=Sym., Noise ratio=50%2026.05 | 95.6 | — | — | — | — | — | |
| PSSCLNoise mode=Sym., Noise ratio=50%2026.05 | 95.6 | — | — | — | — | — | |
| SpikformerParam. (M)=9.32, T=42026.05 | 95.51 | — | — | — | — | — | |
| SpikformerSOPs (G)=66.88, Power (mJ)=5.152026.05 | 95.51 | — | — | — | — | — | |
| LEMONADEGeneration=Full code, Training Protocol=LLM-Based NAS (frozen LLMs, 1-epoch proxy)2026.05 | 95.5 | — | — | — | — | — | |
| ProMixNoise mode=Sym., Noise ratio=20%2026.05 | 95.5 | — | — | — | — | — | |
| CleanBackbone=ResNet-18, Defense Strategy=CutMix2026.06 | 95.5 | — | — | — | — | — | |
| Manifold MixupBackbone=PARN18, Batch Size=128, Optimizer=SGD with momentum 0.92025.09 | 95.43 | — | — | — | — | — | |
| ParBaLS EPIGAL Setup (Initial + Rounds x Batch Size)=100 + 10 x 202025.10 | 95.4 | — | — | — | — | — | |
| CleanBackbone=VGG-11, Filtering Setting=Unfiltered2026.06 | 95.29 | — | — | — | — | — | |
| S^2SAMBackbone=ResNet-18, Pruning Ratio=Dense, WCR=w/o2025.11 | 95.23 | — | — | — | — | — | |
| PSSCLNoise mode=Sym., Noise ratio=80%2026.05 | 95.2 | — | — | — | — | — | |
| SAMBackbone=ResNet-18, Pruning Ratio=Dense, WCR=w/o2025.11 | 95.19 | — | — | — | — | — | |
| CleanBackbone=DenseNet-121, Filtering Setting=Unfiltered2026.06 | 95.12 | — | — | — | — | — | |
| LongReMixNoise mode=Sym., Noise ratio=50%2026.05 | 95.1 | — | — | — | — | — | |
| CleanBackbone=ResNet-18, Defense Strategy=Cutout2026.06 | 95.1 | — | — | — | — | — | |
| CleanBackbone=ResNet-18, Defense Strategy=Mixup2026.06 | 95.01 | — | — | — | — | — | |
| ParBaLS-MAP EPIGAL Setup (Initial + Rounds x Batch Size)=100 + 10 x 202025.10 | 94.93 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Entropy, Label Budget=50%, Backbone=ResNet-1642026.03 | 94.9 | — | — | — | — | — | |
| DivideMixNoise mode=Sym., Noise ratio=50%2026.05 | 94.9 | — | — | — | — | — | |
| GALAXYAL Setup (Initial + Rounds x Batch Size)=100 + 10 x 202025.10 | 94.8 | — | — | — | — | — | |
| Real-ValuedNetwork=ResNet-18, W/A=32/322026.05 | 94.8 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.7, Selection Metric=Least Conf., Label Budget=50%, Backbone=ResNet-1642026.03 | 94.69 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.8, Selection Metric=Least Conf., Label Budget=50%, Backbone=ResNet-1642026.03 | 94.69 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.6, Selection Metric=Least Conf., Label Budget=50%, Backbone=ResNet-1642026.03 | 94.66 | — | — | — | — | — | |
| CleanBackbone=ResNet-18, Defense Strategy=w/o2026.06 | 94.66 | — | — | — | — | — | |
| CleanBackbone=ResNet-18, Filtering Setting=Unfiltered2026.06 | 94.66 | — | — | — | — | — | |
| Baseline ALSelection Metric=Least Conf., Label Budget=50%, Backbone=ResNet-1642026.03 | 94.63 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Least Conf., Label Budget=50%, Backbone=ResNet-1642026.03 | 94.63 | — | — | — | — | — | |
| ProMixNoise mode=Asym., Noise ratio=40%2026.05 | 94.5 | — | — | — | — | — | |
| Baseline ALSelection Metric=Entropy, Label Budget=50%, Backbone=ResNet-1642026.03 | 94.49 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Least Conf., Label Budget=40%, Backbone=ResNet-1642026.03 | 94.41 | — | — | — | — | — | |
| Baseline ALSelection Metric=Greedy k, Label Budget=50%, Backbone=ResNet-1642026.03 | 94.4 | — | — | — | — | — | |
| HRPNoise mode=Sym., Noise ratio=80%2026.05 | 94.4 | — | — | — | — | — | |
| CleanBackbone=ResNet-50, Filtering Setting=Unfiltered2026.06 | 94.37 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Entropy, Label Budget=40%, Backbone=ResNet-1642026.03 | 94.32 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.6, Selection Metric=Least Conf., Label Budget=40%, Backbone=ResNet-1642026.03 | 94.27 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.5, Selection Metric=Greedy k, Label Budget=50%, Backbone=ResNet-1642026.03 | 94.22 | — | — | — | — | — | |
| CrAMBackbone=ResNet-18, Pruning Ratio=Dense, WCR=w/o2025.11 | 94.19 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.7, Selection Metric=Least Conf., Label Budget=40%, Backbone=ResNet-1642026.03 | 94.1 | — | — | — | — | — | |
| Real-ValuedNetwork=VGG-Small, W/A=32/322026.05 | 94.1 | — | — | — | — | — | |
| UNICONNoise mode=Asym., Noise ratio=40%2026.05 | 94.1 | — | — | — | — | — | |
| ProMixNoise mode=Sym., Noise ratio=50%2026.05 | 94.1 | — | — | — | — | — | |
| Baseline ALSelection Metric=Least Conf., Label Budget=40%, Backbone=ResNet-1642026.03 | 94.09 | — | — | — | — | — | |
| PruneFusePruning ratio (p)=0.8, Selection Metric=Least Conf., Label Budget=40%, Backbone=ResNet-1642026.03 | 94.08 | — | — | — | — | — | |
| SAMBackbone=ResNet-18, Pruning Ratio=Dense, WCR=w2025.11 | 94.08 | — | — | — | — | — | |
| Baseline ALSelection Metric=Entropy, Label Budget=40%, Backbone=ResNet-1642026.03 | 94.06 | — | — | — | — | — | |
| RepLBackbone=ResNet-110, GPU Memory (GB)=7.62 (↓18.2%), Training Time (s/epoch)=20.93 (↓20.1%)2026.05 | 94.01 | — | — | — | — | — | |
| UNICONNoise mode=Sym., Noise ratio=20%2026.05 | 94 | — | — | — | — | — | |
| Our workModel Architecture=Hopfield-Resnet132025.09 | 93.92 | — | — | — | — | — | |
| SAMBackbone=ResNet-18, Pruning Ratio=92% Pruned, WCR=w2025.11 | 93.91 | — | — | — | — | — | |
| RRLNoise mode=Sym., Noise ratio=50%2026.05 | 93.9 | — | — | — | — | — | |
| UNICONNoise mode=Sym., Noise ratio=80%2026.05 | 93.9 | — | — | — | — | — | |
| PlainPre-training=Pre-Cifar100, Differential Privacy Status=Non-DP, Training Epochs=1002026.04 | 93.89 | — | — | — | — | — | |
| CrAMBackbone=ResNet-18, Pruning Ratio=Dense, WCR=w2025.11 | 93.86 | — | — | — | — | — | |
| ERMBackbone=PARN18, Batch Size=128, Optimizer=SGD with momentum 0.92025.09 | 93.8 | — | — | — | — | — | |
| LongReMixNoise mode=Sym., Noise ratio=80%2026.05 | 93.8 | — | — | — | — | — | |
| SAMBackbone=ResNet-18, Pruning Ratio=94% Pruned, WCR=w2025.11 | 93.77 | — | — | — | — | — | |
| GEM-FI2026.05 | 93.75 | — | — | — | — | — |