Image Classification on CIFAR-100 (test) (Accuracy and Performance Profile)
81.6Accuracy (%)MAC
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| MACModel=WideResNet-28-10, Epoch=2002025.06 | 81.6 | — | — | — | |
| EVAModel=WideResNet-28-10, Epoch=2002025.06 | 81.6 | — | — | — | |
| SGDModel=WideResNet-28-10, Epoch=2002025.06 | 81.5 | — | — | — | |
| KFACModel=WideResNet-28-10, Epoch=2002025.06 | 81.5 | — | — | — | |
| KFACModel=WideResNet-28-10, Epoch=1002025.06 | 81 | — | — | — | |
| FOOFModel=DenseNet-121, Epoch=2002025.06 | 81 | — | — | — | |
| FOOFModel=WideResNet-28-10, Epoch=2002025.06 | 81 | — | — | — | |
| EVAModel=WideResNet-28-10, Epoch=1002025.06 | 81 | — | — | — | |
| MACModel=WideResNet-28-10, Epoch=1002025.06 | 80.9 | — | — | — | |
| FOOFModel=DenseNet-121, Epoch=1002025.06 | 80.9 | — | — | — | |
| FOOFModel=WideResNet-28-10, Epoch=1002025.06 | 80.8 | — | — | — | |
| SGDModel=WideResNet-28-10, Epoch=1002025.06 | 80.7 | — | — | — | |
| MACModel=DenseNet-121, Epoch=2002025.06 | 80.6 | — | — | — | |
| MACModel=DenseNet-121, Epoch=1002025.06 | 80.5 | — | — | — | |
| ADAMWModel=WideResNet-28-10, Epoch=2002025.06 | 80.2 | — | — | — | |
| KFACModel=DenseNet-121, Epoch=2002025.06 | 80.1 | — | — | — | |
| FOOFModel=WideResNet-28-10, Epoch=502025.06 | 80 | — | — | — | |
| SGDModel=DenseNet-121, Epoch=2002025.06 | 79.9 | — | — | — | |
| EVAModel=DenseNet-121, Epoch=2002025.06 | 79.9 | — | — | — | |
| LNGDModel=DenseNet-121, Epoch=1002025.06 | 79.9 | — | — | — | |
| FOOFModel=DenseNet-121, Epoch=502025.06 | 79.8 | — | — | — | |
| LNGDModel=WideResNet-28-10, Epoch=1002025.06 | 79.8 | — | — | — | |
| ADAMWModel=WideResNet-28-10, Epoch=1002025.06 | 79.7 | — | — | — | |
| KFACModel=DenseNet-121, Epoch=1002025.06 | 79.7 | — | — | — | |
| LNGDModel=DenseNet-121, Epoch=2002025.06 | 79.7 | — | — | — | |
| SGDModel=DenseNet-121, Epoch=1002025.06 | 79.6 | — | — | — | |
| Prune (30%)Backbone=WRN-28-102026.04 | 79.41 | 97.56 | 3.43 | 1 | |
| MACModel=WideResNet-28-10, Epoch=502025.06 | 79.4 | — | — | — | |
| EVAModel=DenseNet-121, Epoch=1002025.06 | 79.4 | — | — | — | |
| LNGDModel=WideResNet-28-10, Epoch=2002025.06 | 79.4 | — | — | — | |
| SGDModel=WideResNet-28-10, Epoch=502025.06 | 79.3 | — | — | — | |
| KFACModel=WideResNet-28-10, Epoch=502025.06 | 79.3 | — | — | — | |
| EVAModel=WideResNet-28-10, Epoch=502025.06 | 79.3 | — | — | — | |
| LNGDModel=WideResNet-28-10, Epoch=502025.06 | 79.1 | — | — | — | |
| Prune (50%)Backbone=WRN-28-102026.04 | 79.07 | 69.69 | 3.36 | 1.02 | |
| ADAMWModel=DenseNet-121, Epoch=2002025.06 | 79 | — | — | — | |
| MACModel=DenseNet-121, Epoch=502025.06 | 78.9 | — | — | — | |
| LNGDModel=DenseNet-121, Epoch=502025.06 | 78.8 | — | — | — | |
| Hybrid (Prune50% → QAT → KD)Backbone=WRN-28-102026.04 | 78.69 | 17.41 | 1.42 | 2.41 | |
| ADAMWModel=DenseNet-121, Epoch=1002025.06 | 78.5 | — | — | — | |
| Hybrid (Prune50% → KD → QAT)Backbone=WRN-28-102026.04 | 78.2 | 17.41 | 1.41 | 2.43 | |
| SGDModel=DenseNet-121, Epoch=502025.06 | 78.2 | — | — | — | |
| ADAMWModel=WideResNet-28-10, Epoch=502025.06 | 78.2 | — | — | — | |
| KFACModel=DenseNet-121, Epoch=502025.06 | 78.1 | — | — | — | |
| EVAModel=DenseNet-121, Epoch=502025.06 | 77.9 | — | — | — | |
| Hybrid (QAT → Prune50% → KD)Backbone=WRN-28-102026.04 | 77.4 | 17.41 | 1.43 | 2.39 | |
| ADAMWModel=DenseNet-121, Epoch=502025.06 | 77.3 | — | — | — | |
| BaselineBackbone=WRN-28-102026.04 | 76.03 | 139.38 | 3.42 | 1 | |
| FOOFModel=ResNet-110, Epoch=2002025.06 | 76 | — | — | — | |
| Hybrid (QAT → KD → Prune50%)Backbone=WRN-28-102026.04 | 75.1 | 17.41 | 1.42 | 2.41 | |
| FOOFModel=ResNet-110, Epoch=1002025.06 | 75.1 | — | — | — | |
| MACModel=ResNet-110, Epoch=2002025.06 | 75 | — | — | — | |
| EVAModel=ResNet-110, Epoch=2002025.06 | 74.7 | — | — | — | |
| LNGDModel=ResNet-110, Epoch=2002025.06 | 74.5 | — | — | — | |
| MACModel=ResNet-110, Epoch=1002025.06 | 74.2 | — | — | — | |
| KFACModel=ResNet-110, Epoch=2002025.06 | 74.2 | — | — | — | |
| ADAMWModel=ResNet-110, Epoch=2002025.06 | 73.7 | — | — | — | |
| FOOFModel=ResNet-110, Epoch=502025.06 | 73.6 | — | — | — | |
| EVAModel=ResNet-110, Epoch=1002025.06 | 73.6 | — | — | — | |
| SGDModel=ResNet-110, Epoch=2002025.06 | 73.5 | — | — | — | |
| ADAMWModel=ResNet-110, Epoch=1002025.06 | 73.4 | — | — | — | |
| KFACModel=ResNet-110, Epoch=1002025.06 | 73.2 | — | — | — | |
| LNGDModel=ResNet-110, Epoch=1002025.06 | 73.1 | — | — | — | |
| MACModel=ResNet-110, Epoch=502025.06 | 72.8 | — | — | — | |
| SGDModel=ResNet-110, Epoch=1002025.06 | 72.5 | — | — | — | |
| KDBackbone=WRN-28-102026.04 | 72.06 | 139.38 | 3.34 | 1.02 | |
| EVAModel=ResNet-110, Epoch=502025.06 | 71.9 | — | — | — | |
| LNGDModel=ResNet-110, Epoch=502025.06 | 71.7 | — | — | — | |
| ADAMWModel=ResNet-110, Epoch=502025.06 | 71.6 | — | — | — | |
| KFACModel=ResNet-110, Epoch=502025.06 | 71.5 | — | — | — | |
| StoMPP + STEArchitecture=R18, Quantization Type=BWN2026.06 | 71.3 | — | — | — | |
| QAT (INT8)Backbone=WRN-28-102026.04 | 71.29 | 35.81 | 1.42 | 2.41 | |
| SGDModel=ResNet-110, Epoch=502025.06 | 71.1 | — | — | — | |
| StoMPPArchitecture=R18, Quantization Type=BWN2026.06 | 69.5 | — | — | — | |
| StoMPPArchitecture=R50, Quantization Type=BWN2026.06 | 69 | — | — | — | |
| StoMPP + STEArchitecture=R34, Quantization Type=BWN2026.06 | 68.9 | — | — | — | |
| StoMPP + STEArchitecture=R50, Quantization Type=BWN2026.06 | 68.3 | — | — | — | |
| StoMPPArchitecture=MobileNetV2, Quantization Type=BWN2026.06 | 67.7 | — | — | — | |
| StoMPP + STEArchitecture=MobileNetV2, Quantization Type=BWN2026.06 | 67.7 | — | — | — | |
| StoMPPArchitecture=R34, Quantization Type=BWN2026.06 | 66.3 | — | — | — | |
| LeJEPA(L SIGReg)Pre-train dataset=Inet100, Pre-train epochs=800, Evaluation protocol=frozen-backbone linear-probe, Backbone features=concat of the last two CLS tokens, Probe epochs=50, Probe optimizer=AdamW with cosine annealing2026.05 | 65.33 | — | — | — | |
| STEArchitecture=MobileNetV2, Quantization Type=BWN2026.06 | 65.3 | — | — | — | |
| STEArchitecture=R34, Quantization Type=BWN2026.06 | 64.9 | — | — | — | |
| STEArchitecture=R18, Quantization Type=BWN2026.06 | 64.6 | — | — | — | |
| UR-JEPA(L CGLT)Pre-train dataset=Inet100, Pre-train epochs=800, Evaluation protocol=frozen-backbone linear-probe, Backbone features=concat of the last two CLS tokens, Probe epochs=50, Probe optimizer=AdamW with cosine annealing2026.05 | 64.48 | — | — | — | |
| STEArchitecture=R50, Quantization Type=BWN2026.06 | 64.3 | — | — | — | |
| StoMPP + STEArchitecture=R18, Quantization Type=BNN2026.06 | 58 | — | — | — | |
| StoMPPArchitecture=R18, Quantization Type=BNN2026.06 | 53.8 | — | — | — | |
| STEArchitecture=R18, Quantization Type=BNN2026.06 | 49.1 | — | — | — | |
| StoMPP + STEArchitecture=MobileNetV2, Quantization Type=BNN2026.06 | 48 | — | — | — | |
| StoMPP + STEArchitecture=R34, Quantization Type=BNN2026.06 | 47.8 | — | — | — | |
| StoMPP + STEArchitecture=R50, Quantization Type=BNN2026.06 | 46.5 | — | — | — | |
| StoMPPArchitecture=MobileNetV2, Quantization Type=BNN2026.06 | 42 | — | — | — | |
| STEArchitecture=MobileNetV2, Quantization Type=BNN2026.06 | 40.5 | — | — | — | |
| StoMPPArchitecture=R50, Quantization Type=BNN2026.06 | 40.2 | — | — | — | |
| StoMPPArchitecture=R34, Quantization Type=BNN2026.06 | 39.8 | — | — | — | |
| STEArchitecture=R34, Quantization Type=BNN2026.06 | 33.7 | — | — | — | |
| STEArchitecture=R50, Quantization Type=BNN2026.06 | 26.7 | — | — | — |