Image Classification on ImageNet (10% labels)
84.3Top-1 AccSemi-ViT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Semi-ViTArchitecture=ViT-Huge, Param=632M2022.08 | 84.3 | — | — | — | — | |
| Semi-ViTArchitecture=ViT-Large, Param=307M2022.08 | 83.3 | — | — | — | — | |
| SimCLRv2+KDArchitecture=RN152 (3x+SK), Param=794M2022.08 | 80.9 | — | — | — | — | |
| Semi-ViTArchitecture=ViT-Base, Param=86M2022.08 | 79.7 | — | — | — | — | |
| CLOPBackbone=ResNet-50 (4x)2025.11 | 79.1 | 92.7 | — | — | — | |
| Semi-ViTArchitecture=ViT-Small, Param=22M2022.08 | 77.1 | — | — | — | — | |
| PAWSBackbone=ResNet-50, Epochs=300, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 75.5 | — | — | — | — | |
| PAWSArchitecture=ResNet-50, Param=26M2022.08 | 75.5 | — | — | — | — | |
| SemiFormerArchitecture=ViT-S+Conv, Param=42M2022.08 | 75.5 | — | — | — | — | |
| SequenceMatchSelf-supervised pre-training=NONE, Params (train/test)=25.6M/25.6M, Epochs=~ 4002023.10 | 75.2 | 91.9 | — | — | — | |
| CCSSLBackbone=ResNet-50 (4x)2025.11 | 75.1 | 92.3 | — | — | — | |
| PAWSBackbone=ResNet-50, Epochs=200, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 75 | — | — | — | — | |
| SimMatch-V2Backbone=ResNet-50 (4x)2025.11 | 74.8 | 91.7 | — | — | — | |
| MoProPretrain dataset=WebVision-V1.0, #Pretrain epochs=902020.09 | 74.8 | 92.4 | — | — | — | |
| CoMatch + SemiRewardBackbone=ViT, Variant=+SR2023.10 | 74.5 | — | — | — | — | |
| SimMatchSelf-supervised Pre-training=None, Epochs=~400, Parameters (train/test)=30.0M / 25.6M2022.03 | 74.4 | 91.6 | — | — | — | |
| SimMatchSelf-supervised pre-training=NONE, Params (train/test)=30.0M/25.6M, Epochs=~ 4002023.10 | 74.4 | 91.6 | — | — | — | |
| SimMatchBackbone=ViT, Variant=Base2023.10 | 74.4 | — | — | — | — | |
| FixMatch-EMANArch=res502021.01 | 74 | 90.9 | — | — | — | |
| FixMatch-EMANSelf-supervised Pre-training=MoCo-EMAN, Epochs=~1100, Parameters (train/test)=30.0M / 25.6M2022.03 | 74 | 90.9 | — | — | — | |
| EMANArchitecture=ResNet-50, Param=26M2022.08 | 74 | — | — | — | — | |
| FixMatch-EMANSelf-supervised pre-training=MoCo-EMAN [6], Params (train/test)=30.0M/25.6M, Epochs=~ 11002023.10 | 74 | 90.9 | — | — | — | |
| SimMatchBackbone=ResNet-50 (4x)2025.11 | 74 | 93 | — | — | — | |
| MPLBackbone=ResNet-50, Epochs=800, Evaluation Protocol=Label Propagation2021.04 | 73.9 | — | — | — | — | |
| PAWSBackbone=ResNet-50, Epochs=100, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 73.9 | — | — | — | — | |
| MPLArchitecture=ResNet-50, Param=26M2022.08 | 73.9 | — | — | — | — | |
| CowMixArchitecture=ResNet-152, Param=60M2022.08 | 73.9 | — | — | — | — | |
| CoMatchSelf-supervised Pre-training=MoCo V2, Epochs=~1200, Parameters (train/test)=30.0M / 25.6M2022.03 | 73.7 | 91.4 | — | — | — | |
| CoMatchSelf-supervised pre-training=MoCo V2 [9], Params (train/test)=30.0M/25.6M, Epochs=~ 12002023.10 | 73.7 | 91.4 | — | — | — | |
| MoProPretrain dataset=WebVision-V2.0, #Pretrain epochs=902020.09 | 73.7 | 92.1 | — | — | — | |
| CoMatchSelf-supervised Pre-training=None, Epochs=~400, Parameters (train/test)=30.0M / 25.6M2022.03 | 73.6 | 91.6 | — | — | — | |
| CoMatchSelf-supervised pre-training=NONE, Params (train/test)=30.0M/25.6M, Epochs=~ 4002023.10 | 73.6 | 91.6 | — | — | — | |
| CoMatchBackbone=ViT, Variant=Base2023.10 | 73.6 | — | — | — | — | |
| CEPretrain dataset=WebVision-V1.0, #Pretrain epochs=902020.09 | 73.5 | 91.7 | — | — | — | |
| MoProPretrain dataset=WebVision-V0.5, #Pretrain epochs=902020.09 | 73.3 | 91.7 | — | — | — | |
| S4LArchitecture=ResNet-50 (4x), Param=375M2022.08 | 73.2 | — | — | — | — | |
| PAWS-NNBackbone=ResNet-50, Epochs=300, Evaluation Protocol=Non-parametric classification (No fine-tuning)2021.04 | 73.1 | — | — | — | — | |
| CEPretrain dataset=WebVision-V2.0, #Pretrain epochs=902020.09 | 72.9 | 91.4 | — | — | — | |
| FixMatch + SemiRewardBackbone=ViT, Variant=+SR2023.10 | 72.8 | — | — | — | — | |
| CEPretrain dataset=WebVision-V0.5, #Pretrain epochs=902020.09 | 72.4 | 90.9 | — | — | — | |
| DINOArchitecture=ViT-Small, Param=22M2022.08 | 72.2 | — | — | — | — | |
| SsCLBackbone=ResNet-50 (4x)2025.11 | 72.1 | 90.9 | — | — | — | |
| Fine-tuneSelf-supervised Pre-training=WCL, Epochs=~800, Parameters (train/test)=34.2M / 25.6M2022.03 | 72 | 91.2 | — | — | — | |
| FixMatchBackbone=ResNet-50 (4x)2025.11 | 72 | 88.6 | — | — | — | |
| PAWS-NNBackbone=ResNet-50, Epochs=200, Evaluation Protocol=Non-parametric classification (No fine-tuning)2021.04 | 71.9 | — | — | — | — | |
| TWIST2021.12 | 71.7 | 91 | — | — | — | |
| FixMatchBackbone=ViT, Variant=Base2023.10 | 71.6 | — | — | — | — | |
| FixMatchArch=res502021.01 | 71.5 | 89.1 | — | — | — | |
| FixMatchBackbone=ResNet-50, Epochs=300, Evaluation Protocol=Label Propagation2021.04 | 71.5 | — | — | — | — | |
| FixMatchSelf-supervised Pre-training=None, Epochs=~300, Parameters (train/test)=25.6M / 25.6M2022.03 | 71.5 | 89.1 | — | — | — | |
| UniGrad2021.12 | 71.5 | 90.6 | — | — | — | |
| FixMatchArchitecture=ResNet-50, Param=26M2022.08 | 71.5 | — | — | — | — | |
| FixMatchSelf-supervised pre-training=NONE, Params (train/test)=25.6M/25.6M, Epochs=~ 3002023.10 | 71.5 | 89.1 | — | — | — | |
| RényiCLBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 71.2 | 90.3 | — | — | — | |
| PAWS-NNBackbone=ResNet-50, Epochs=100, Evaluation Protocol=Non-parametric classification (No fine-tuning)2021.04 | 71 | — | — | — | — | |
| SwAV+CTBackbone=ResNet-50, Epochs=400, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 70.8 | — | — | — | — | |
| SimCLR-v2Arch=res502021.01 | 70.5 | — | — | — | — | |
| SimCLRv2 (+Self.Dist.)Backbone=ResNet-50, Epochs=1200, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 70.5 | — | — | — | — | |
| C-BYOLBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 70.5 | 90 | — | — | — | |
| SupConBackbone=ResNet-50 (4x)2025.11 | 70.4 | 87.4 | — | — | — | |
| SwAVBackbone=ResNet-50, Epochs=800, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 70.2 | — | — | — | — | |
| Fine-tuneSelf-supervised Pre-training=SwAV, Epochs=~800, Parameters (train/test)=30.4M / 25.6M2022.03 | 70.2 | 89.9 | — | — | — | |
| SwAVBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 70.2 | 89.9 | — | — | — | |
| SwAVPretrain dataset=ImageNet, #Pretrain epochs=8002020.09 | 70.2 | 89.9 | — | — | — | |
| NNCLRBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 69.8 | 89.3 | — | — | — | |
| Barlow Twins2021.12 | 69.7 | 89.3 | — | — | — | |
| Barlow TwinsBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 69.7 | 89.3 | — | — | — | |
| UDAArch=res502021.01 | 68.8 | 88.5 | — | — | — | |
| BYOLBackbone=ResNet-50, Epochs=1000, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 68.8 | — | — | — | — | |
| VAT+EntMin.Self-supervised Pre-training=None, Parameters (train/test)=25.6M / 25.6M2022.03 | 68.8 | 88.5 | — | — | — | |
| UDASelf-supervised Pre-training=None, Parameters (train/test)=25.6M / 25.6M2022.03 | 68.8 | 88.5 | — | — | — | |
| Fine-tuneSelf-supervised Pre-training=BYOL, Epochs=~1000, Parameters (train/test)=37.1M / 25.6M2022.03 | 68.8 | 89 | — | — | — | |
| BYOL2021.12 | 68.8 | 89 | — | — | — | |
| UDAArchitecture=ResNet-50, Param=26M2022.08 | 68.8 | — | — | — | — | |
| BYOLBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 68.8 | 89 | — | — | — | |
| BYOLPretrain dataset=ImageNet, #Pretrain epochs=10002020.09 | 68.8 | 89 | — | — | — | |
| UDA (RandAugment)SSL=true, Backbone=ResNet-50, Image size=2242019.04 | 68.78 | 88.8 | — | — | — | |
| SimCLRv2Backbone=ResNet-50, Epochs=800, Evaluation Protocol=Representation Learning (Fine-tuning)2021.04 | 68.4 | — | — | — | — | |
| Fine-tuneSelf-supervised Pre-training=SimCLR V2, Epochs=~800, Parameters (train/test)=34.2M / 25.6M2022.03 | 68.4 | 89.2 | — | — | — | |
| DINO2021.12 | 68.2 | 89.1 | — | — | — | |
| UDABackbone=ResNet-50, Epochs=800, Evaluation Protocol=Label Propagation2021.04 | 68.1 | — | — | — | — | |
| DINO + SDMPModel=ViT-S, Parameters=21M, Pre-training epochs=3002022.06 | 68 | — | — | — | — | |
| MoCo v3 + SDMPModel=ViT-S, Parameters=21M, Pre-training epochs=3002022.06 | 67.4 | — | — | — | — | |
| DINOModel=ViT-S, Parameters=21M, Pre-training epochs=3002022.06 | 67.2 | — | — | — | — | |
| MoCo v3Model=ViT-S, Parameters=21M, Pre-training epochs=3002022.06 | 66.7 | — | — | — | — | |
| Fine-tuneSelf-supervised Pre-training=MoCo V2, Epochs=~800, Parameters (train/test)=30.0M / 25.6M2022.03 | 66.1 | 87.9 | — | — | — | |
| Fine-tuneSelf-supervised Pre-training=SimCLR, Epochs=~1000, Parameters (train/test)=30.0M / 25.6M2022.03 | 65.6 | 87.8 | — | — | — | |
| SimCLR2021.12 | 65.6 | 87.8 | — | — | — | |
| SimCLRBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 65.6 | 87.8 | — | — | — | |
| SimCLRPretrain dataset=ImageNet, #Pretrain epochs=10002020.09 | 65.6 | 87.8 | — | — | — | |
| PCLPretrain dataset=ImageNet, #Pretrain epochs=2002020.09 | 62.9 | 85.6 | — | — | — | |
| ResNet-50 w. RandAugmentSSL=false, Backbone=ResNet-50, Image size=2242019.04 | 58.84 | 80.56 | — | — | — | |
| SupervisedArch=res502021.01 | 56.4 | 80.4 | — | — | — | |
| Supervised2021.12 | 56.4 | 80.4 | — | — | — | |
| SupervisedBackbone=ResNet-50, Evaluation protocol=fine-tuning2022.08 | 56.4 | 80.4 | — | — | — | |
| Random init.Pretrain dataset=None, #Pretrain epochs=None2020.09 | 56.4 | 80.4 | — | — | — | |
| ResNet-50SSL=false, Backbone=ResNet-50, Image size=2242019.04 | 55.09 | 77.26 | — | — | — | |
| S4L-RotationSelf-supervised Pre-training=None, Epochs=~200, Parameters (train/test)=25.6M / 25.6M2022.03 | 53.4 | 83.8 | — | — | — | |
| AMDIMMethod Category=Representation learning, Evaluation Protocol=Fixed2019.05 | — | 85.8 | — | — | — | |
| BigBiGANMethod Category=Representation learning, Evaluation Protocol=Fixed2019.05 | — | 78.8 | — | — | — |