Semi-supervised Image Classification on ImageNet 1k (1%)
77.3Top-1 AccI-JEPA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| I-JEPAArch.=ViT-H/16, Resolution=448x448, Epochs=300, View Data Augmentation=without2023.01 | 77.3 | — | |
| MSN + IbM2Backbone=ViT-B/4, IbM2=true2023.05 | 76.1 | — | |
| MSNArch.=ViT-B/4, Epochs=300, View Data Augmentation=extra2023.01 | 75.7 | — | |
| MSN + IbM2Backbone=ViT-L/7, IbM2=true2023.05 | 75.6 | — | |
| MSNBackbone=ViT-B/4, IbM2=false2023.05 | 75.4 | — | |
| MSNBackbone=ViT-L/7, IbM2=false2023.05 | 74.8 | — | |
| data2vecArch.=ViT-L/16, Epochs=1600, View Data Augmentation=without2023.01 | 73.3 | — | |
| I-JEPAArch.=ViT-H/14, Epochs=300, View Data Augmentation=without2023.01 | 73.3 | — | |
| MAEArch.=ViT-H/14, Epochs=1600, View Data Augmentation=without2023.01 | 71.5 | — | |
| BYOLArch.=R200 (x2), Param.=250M, Repr.=40962021.05 | 71.2 | 89.5 | |
| BYOLArch.=RN200 (2x), Epochs=800, View Data Augmentation=extra2023.01 | 71.2 | — | |
| DINO + IbM2Backbone=ViT-S/8, IbM2=true2023.05 | 70.6 | — | |
| SimCLR v2Arch.=RN151 (2x), Epochs=800, View Data Augmentation=extra2023.01 | 70.2 | — | |
| DINOBackbone=ViT-S/8, IbM2=false2023.05 | 70.1 | — | |
| DINOArch.=ViT-B/8, Epochs=300, View Data Augmentation=extra2023.01 | 70 | — | |
| iBOTArch.=ViT-B/16, Epochs=400, View Data Augmentation=extra2023.01 | 69.7 | — | |
| I-JEPAArch.=ViT-L/16, Epochs=600, View Data Augmentation=without2023.01 | 69.4 | — | |
| BYOLArch.=R50 (x4), Param.=375M, Repr.=81922021.05 | 69.1 | 87.9 | |
| VICRegArch.=R200 (x2), Param.=250M, Repr.=40962021.05 | 68.8 | 88.2 | |
| MSN + IbM2Backbone=ViT-S/16, IbM2=true2023.05 | 68 | — | |
| MSNBackbone=ViT-S/16, IbM2=false2023.05 | 67.3 | — | |
| MAEArch.=ViT-L/16, Epochs=1600, View Data Augmentation=without2023.01 | 67.1 | — | |
| DINO + IbM2Backbone=ViT-S/16, IbM2=true2023.05 | 65.2 | — | |
| DINOBackbone=ViT-S/16, IbM2=false2023.05 | 64.7 | — | |
| FNCBackbone=ResNet-50, Learning Type=Representation Learning / Contrastive2020.11 | 63.7 | 85.3 | |
| SimCLRArch.=R50 (x4), Param.=375M, Repr.=81922021.05 | 63 | 85.8 | |
| VICRegArch.=R50 (x2), Param.=93M, Repr.=40962021.05 | 62.6 | 84.5 | |
| BYOLArch.=R50 (x2), Param.=93M, Repr.=40962021.05 | 62.2 | 84.1 | |
| C-BYOLBackbone=ResNet-502022.08 | 60.6 | 83.4 | |
| MoCov3 + IbM2Backbone=ViT-S/16, IbM2=true2023.05 | 59.1 | — | |
| SimCLRArch.=R50 (x2), Param.=93M, Repr.=40962021.05 | 58.5 | 83 | |
| MoCov3Backbone=ViT-S/16, IbM2=false2023.05 | 58.1 | — | |
| SimCLR v2Backbone=ResNet-50, Learning Type=Representation Learning / Contrastive2020.11 | 57.9 | 82.5 | |
| BYOL + IbM2Backbone=ResNet50, IbM2=true2023.05 | 56.9 | — | |
| NNCLRBackbone=ResNet-502022.08 | 56.4 | 80.7 | |
| RényiCLBackbone=ResNet-502022.08 | 56.4 | 80.6 | |
| BYOLBackbone=ResNet50, IbM2=false2023.05 | 55.7 | — | |
| Barlow TwinsBackbone=ResNet-502022.08 | 55 | 79.2 | |
| SwAVBackbone=ResNet-502020.11 | 53.9 | 78.5 | |
| SwAVBackbone=ResNet-502022.08 | 53.9 | 78.5 | |
| BYOLBackbone=ResNet-502020.11 | 53.2 | 78.4 | |
| BYOLBackbone=ResNet-502022.08 | 53.2 | 78.4 | |
| SimCLR + IbM2Backbone=ResNet50, IbM2=true2023.05 | 51.3 | — | |
| SimCLRBackbone=ResNet50, IbM2=false2023.05 | 50.5 | — | |
| SimCLR v1Backbone=ResNet-50, Learning Type=Representation Learning / Contrastive2020.11 | 48.3 | 75.5 | |
| SimCLRBackbone=ResNet-502022.08 | 48.3 | 75.5 | |
| PIRLBackbone=ResNet-50, Learning Type=Representation Learning / Contrastive2020.11 | 30.7 | 60.4 | |
| SupervisedBackbone=ResNet-502020.11 | 25.4 | 56.4 | |
| SupervisedBackbone=ResNet-502022.08 | 25.4 | 48.4 | |
| FixMatchBackbone=ResNet-50, Learning Type=Semi-supervised2020.11 | — | 71.5 | |
| PCLBackbone=ResNet-50, Learning Type=Representation Learning / Contrastive2020.11 | — | 75.6 | |
| UDABackbone=ResNet-50, Learning Type=Semi-supervised2020.11 | — | 68.8 |