Image Classification on CIFAR-100 (test) (Accuracy and Throughput)
89.84AccuracySGD
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SGDBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w/o2025.11 | 89.84 | — | |
| SGDBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w2025.11 | 89.75 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w/o2025.11 | 88.71 | — | |
| SGDBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w2025.11 | 88.58 | — | |
| DP-GRAPEepsilon=8, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 88.1 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w/o2025.11 | 88.06 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w2025.11 | 87.97 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w2025.11 | 87.9 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w2025.11 | 87.82 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w2025.11 | 87.82 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w2025.11 | 87.82 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=Dense, WCR=w/o2025.11 | 87.36 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w2025.11 | 87.08 | — | |
| DP-GRAPEepsilon=4, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 86.9 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w2025.11 | 86.8 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w2025.11 | 86.7 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w/o2025.11 | 86.33 | — | |
| SGDBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w/o2025.11 | 86.3 | — | |
| SGDBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w2025.11 | 85.66 | — | |
| Naïve DP-GaLoreepsilon=8, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 85.5 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w2025.11 | 85.47 | — | |
| DP-GRAPEepsilon=2, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 85.4 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=80% Pruned, WCR=w2025.11 | 85.25 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=80% Pruned, WCR=w2025.11 | 84.95 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w/o2025.11 | 84.83 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=60% Pruned, WCR=w/o2025.11 | 84.82 | — | |
| DINOv3-S/16 (LOES)Model=DINOv3-S/16, Selection Strategy=LOES2026.05 | 84.73 | — | |
| DINOv2-S (LOES)Model=DINOv2-S, Selection Strategy=LOES2026.05 | 84.31 | — | |
| DINOv3-S/16 (Last)Model=DINOv3-S/16, Selection Strategy=Last Layer2026.05 | 84.07 | — | |
| Naïve DP-GaLoreepsilon=4, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 83.8 | — | |
| DeiT-B/16 (LOES)Model=DeiT-B/16, Selection Strategy=LOES2026.05 | 83.75 | — | |
| RW-SAMModel=WideResNet2025.09 | 83.52 | — | |
| SAMModel=WideResNet2025.09 | 83.25 | — | |
| DINOv2-S (Last)Model=DINOv2-S, Selection Strategy=Last Layer2026.05 | 83.2 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w/o2025.11 | 82.8 | — | |
| DeiT-B/16 (Last)Model=DeiT-B/16, Selection Strategy=Last Layer2026.05 | 82.49 | — | |
| S^2SAMBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w/o2025.11 | 81.95 | — | |
| SGDBackbone=WRN-28-10, Sparsity=-2026.03 | 81.63 | 752.3 | |
| SGDModel=WideResNet2025.09 | 81.55 | — | |
| DP-GRAPEepsilon=1, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 81.4 | — | |
| RW-SAMModel=ResNet-502025.09 | 80.83 | — | |
| DP-Adamepsilon=8, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 80.8 | — | |
| CLIP-B32 (LOES)Model=CLIP-B32, Selection Strategy=LOES2026.05 | 80.59 | — | |
| SAMModel=ResNet-502025.09 | 80.31 | — | |
| SAMBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w/o2025.11 | 80.29 | — | |
| GSAMBackbone=WRN-28-10, Sparsity=90%2026.03 | 80.28 | 340.19 | |
| ESAMBackbone=WRN-28-10, Sparsity=90%2026.03 | 80.23 | 305.76 | |
| SAMBackbone=WRN-28-10, Sparsity=90%2026.03 | 80.14 | 354.94 | |
| ZO-SAMBackbone=WRN-28-10, Sparsity=90%2026.03 | 79.91 | 576.01 | |
| SGDModel=ResNet-502025.09 | 79.55 | — | |
| SAMBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w2025.11 | 79.51 | — | |
| ViT-IN21k-B/16 (LOES)Model=ViT-IN21k-B/16, Selection Strategy=LOES2026.05 | 79.44 | — | |
| RW-SAMModel=ResNet-182025.09 | 79.31 | — | |
| LS(k=5)Backbone=WRN-28-10, Sparsity=90%2026.03 | 79.3 | 542.63 | |
| S^2SAMBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w/o2025.11 | 79.3 | — | |
| S^2SAMBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w2025.11 | 79.3 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=80% Pruned, WCR=w2025.11 | 79.29 | — | |
| SGDModel=ResNet-182025.09 | 78.91 | — | |
| SAMModel=ResNet-182025.09 | 78.9 | — | |
| LS(k=10)Backbone=WRN-28-10, Sparsity=90%2026.03 | 78.84 | 593.02 | |
| Naïve DP-GaLoreepsilon=2, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 78.8 | — | |
| SAMBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w/o2025.11 | 78.43 | — | |
| SGDBackbone=ViT-B/32, Pruning Ratio=70% Pruned, WCR=w/o2025.11 | 78.33 | — | |
| SAMBackbone=WideRes-28-10, Pruning Ratio=92% Pruned, WCR=w2025.11 | 77.95 | — | |
| ViT-IN21k-B/16 (Last)Model=ViT-IN21k-B/16, Selection Strategy=Last Layer2026.05 | 77.43 | — | |
| S^2SAMBackbone=WideRes-28-10, Pruning Ratio=92% Pruned, WCR=w2025.11 | 77.33 | — | |
| DP-Adamepsilon=4, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 77.3 | — | |
| SGD (well-tuned)Backbone=ResNet-18, Training Setup=Single-worker2026.04 | 77 | — | |
| CrAMBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w/o2025.11 | 76.29 | — | |
| SignSGD-M + pre-noise (α=0.1)Backbone=ResNet-18, Training Setup=Single-worker, Noise Type=pre-noise, α=0.12026.04 | 76 | — | |
| SGDBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w/o2025.11 | 75.93 | — | |
| CrAMBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w2025.11 | 75.86 | — | |
| SGDBackbone=WideRes-28-10, Pruning Ratio=Dense, WCR=w2025.11 | 75.82 | — | |
| CLIP-B32 (Last)Model=CLIP-B32, Selection Strategy=Last Layer2026.05 | 75.63 | — | |
| SGDBackbone=ResNet-32, Sparsity=-2026.03 | 74.89 | 5,673.95 | |
| SANERNoise type=Asymmetric noise, Noise rate=25%, Backbone=ResNet182024.11 | 74.64 | — | |
| SAMBackbone=WideRes-28-10, Pruning Ratio=94% Pruned, WCR=w2025.11 | 73.97 | — | |
| S^2SAMBackbone=WideRes-28-10, Pruning Ratio=94% Pruned, WCR=w2025.11 | 73.95 | — | |
| SGDBackbone=ViT-B/32, Pruning Ratio=80% Pruned, WCR=w2025.11 | 73.69 | — | |
| OptimalSearch Space=NAS-Bench-2012026.04 | 73.51 | — | |
| SignSGD-M + pre-noise (α=0.5)Backbone=ResNet-18, Training Setup=Single-worker, Noise Type=pre-noise, α=0.52026.04 | 73 | — | |
| SANERNoise type=Dependent noise, Noise rate=25%, Backbone=ResNet182024.11 | 72.93 | — | |
| GSAMBackbone=ResNet-32, Sparsity=90%2026.03 | 72.9 | 2,701 | |
| SANERNoise type=Symmetric noise, Noise rate=25%, Backbone=ResNet182024.11 | 72.9 | — | |
| SAMBackbone=ResNet-32, Sparsity=90%2026.03 | 72.64 | 2,704.84 | |
| G-ICSO-NASSearch Space=NAS-Bench-2012026.04 | 72.63 | — | |
| ESAMBackbone=ResNet-32, Sparsity=90%2026.03 | 72.49 | 2,297.23 | |
| ZO-SAMBackbone=ResNet-32, Sparsity=90%2026.03 | 72.2 | 4,349.53 | |
| LS(k=5)Backbone=ResNet-32, Sparsity=90%2026.03 | 72.01 | 3,980.62 | |
| Hybrid (SignSGD-M → SGD)Backbone=ResNet-18, Training Setup=Single-worker2026.04 | 72 | — | |
| LS(k=10)Backbone=ResNet-32, Sparsity=90%2026.03 | 71.62 | 4,272.22 | |
| SAMNoise type=Asymmetric noise, Noise rate=25%, Backbone=ResNet182024.11 | 71.57 | — | |
| DARTS-Search Space=NAS-Bench-2012026.04 | 71.53 | — | |
| Our workModel Architecture=Hopfield-Resnet132025.09 | 71.05 | — | |
| ResNetSearch Space=NAS-Bench-2012026.04 | 70.86 | — | |
| iDARTSSearch Space=NAS-Bench-2012026.04 | 70.83 | — | |
| Our workModel Architecture=VGG52025.09 | 70.78 | — | |
| Randomtype=baseline, Search Space=NAS-Bench-2012026.04 | 70.65 | — | |
| DP-Adamepsilon=2, Backbone=Vision Transformer, Fine-tuning=true2025.06 | 70.6 | — | |
| CrAMBackbone=ViT-B/32, Pruning Ratio=80% Pruned, WCR=w/o2025.11 | 70.31 | — |