Image Classification on STL-10 (test)
99.71AccuracyViT-L/16 + (Spinal FC & Background)
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ViT-L/16 + (Spinal FC & Background)Architecture=ViT-L/16, Epochs=2, Spinal FC=true, Background Class=true2023.05 | 99.71 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ViT-L/16 + BackgroundArchitecture=ViT-L/16, Epochs=2, Spinal FC=false, Background Class=true2023.05 | 99.63 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ViT-L/16 + Spinal FCArchitecture=ViT-L/16, Epochs=2, Spinal FC=true, Background Class=false2023.05 | 99.58 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ViT-L/16Architecture=ViT-L/16, Epochs=2, Spinal FC=false, Background Class=false2023.05 | 99.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| OTTERBackbone=ViT-B/16, Setting=Zero-shot2024.04 | 98.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIP ViT-B-16Parameters=86.2 M, Evaluation Protocol=linear evaluation2023.03 | 98.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Prior MatchingBackbone=ViT-B/16, Setting=Zero-shot2024.04 | 98.4 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zero-shotBackbone=ViT-B/16, Setting=Zero-shot2024.04 | 98 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ProDAshots=162022.05 | 96.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ProDAshots=82022.05 | 96.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ProDAshots=42022.05 | 95.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoOpshots=162022.05 | 95.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MV-MR ResNet50Parameters=23.5 M, Evaluation Protocol=linear evaluation2023.03 | 95.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoOpshots=82022.05 | 95.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ProDAshots=22022.05 | 95.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ProDAshots=12022.05 | 95.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Linear Probe CLIPshots=162022.05 | 95 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoOpshots=42022.05 | 94.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Zero-Shot CLIPshots=02022.05 | 94.4 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Linear Probe CLIPshots=82022.05 | 94.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoOpshots=22022.05 | 94.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoOpshots=12022.05 | 94.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| AMDIM(1)source=Reported in the first version of [3], evaluation_protocol=finetuning a 1,024-unit MLP on top of a frozen, pretrained encoder2020.08 | 93.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| AMDIM(2)source=Reported in the second version of [3], evaluation_protocol=finetuning a 1,024-unit MLP on top of a frozen, pretrained encoder2020.08 | 93.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RDSSSSL Framework=FlexMatch, Annotation Budget=2502024.09 | 93.16 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DiffAugEvaluation Protocol=Linear probing2023.09 | 92.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RDSSSSL Framework=FreeMatch, Annotation Budget=2502024.09 | 92.22 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Linear Probe CLIPshots=42022.05 | 92.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| YADIMsource=This paper, evaluation_protocol=finetuning a 1,024-unit MLP on top of a frozen, pretrained encoder2020.08 | 92.15 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BYOLEvaluation Protocol=Linear probing2023.09 | 91.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DINOEvaluation Protocol=Linear probing2023.09 | 91.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR + VAEEvaluation Protocol=Linear probing, Augmentation=VAE2023.09 | 91.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| AMDIMsource=Own implementation, evaluation_protocol=finetuning a 1,024-unit MLP on top of a frozen, pretrained encoder2020.08 | 91.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCoV2 + VAEEvaluation Protocol=Linear probing, Augmentation=VAE2023.09 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCoV2 + GANEvaluation Protocol=Linear probing, Augmentation=GAN2023.09 | 91.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCoV2 + MixupEvaluation Protocol=Linear probing, Augmentation=Mixup2023.09 | 90.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR†+IP-IRMTraining Epochs=1000, Evaluation Protocol=Linear, MixUp=true2021.10 | 89.91 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR + GANEvaluation Protocol=Linear probing, Augmentation=GAN2023.09 | 89.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MV-MREvaluation protocol=Linear evaluation on frozen representations2023.03 | 89.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR + MixupEvaluation Protocol=Linear probing, Augmentation=Mixup2023.09 | 89.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimSiamEvaluation Protocol=Linear probing2023.09 | 89.4 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCo-v2 + Patch-based NSMethod Variant=Patch-based NS2021.10 | 89.36 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCoV2Evaluation Protocol=Linear probing2023.09 | 89.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLREvaluation Protocol=Linear probing2023.09 | 89 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ReSSLBackProp=1x, EMA=Yes, Evaluation Protocol=Linear evaluation2021.07 | 88.25 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCo-v22021.10 | 88 | — | — | — | — | — | — | — | — | — | — | — | — | |
| InfoNCE (Lnce)Backbone=ResNet50, Epochs=500, Batchsize=128, Learning rate=0.001, Evaluation protocol=linear evaluation2022.01 | 87.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| α-CL-directBackbone=ResNet50, Epochs=500, Batchsize=128, Learning rate=0.001, Evaluation protocol=linear evaluation2022.01 | 87.85 | — | — | — | — | — | — | — | — | — | — | — | — | |
| HCL+IP-IRMTraining Epochs=400, Evaluation Protocol=Linear2021.10 | 87.81 | — | — | — | — | — | — | — | — | — | — | — | — | |
| HCLTraining Epochs=400, Evaluation Protocol=Linear2021.10 | 87.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimSiamBackProp=2x, EMA=No, Evaluation Protocol=Linear evaluation2021.07 | 87.47 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BYOLBackProp=2x, EMA=Yes, Evaluation Protocol=Linear evaluation2021.07 | 87.45 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR + HardMethod Variant=Hard Negatives2021.10 | 87.42 | — | — | — | — | — | — | — | — | — | — | — | — | |
| α-CL-directBackbone=ResNet50, Epochs=300, Batchsize=128, Learning rate=0.001, Evaluation protocol=linear evaluation2022.01 | 87.17 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Linear Probe CLIPshots=22022.05 | 86.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| RUCEvaluation protocol=Linear evaluation on frozen representations2023.03 | 86.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| InfoNCE (Lnce)Backbone=ResNet50, Epochs=300, Batchsize=128, Learning rate=0.001, Evaluation protocol=linear evaluation2022.01 | 86.57 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CyCLIPzero-shot=true2022.05 | 86.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| MoCoV2BackProp=1x, EMA=Yes, Evaluation Protocol=Linear evaluation2021.07 | 85.88 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR†Training Epochs=1000, Evaluation Protocol=Linear, MixUp=true2021.10 | 85.56 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLRBackProp=2x, EMA=No, Evaluation Protocol=Linear evaluation2021.07 | 85.48 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEConvTraining Augmentation=AugMix2023.10 | 85.46 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DCL+IP-IRMTraining Epochs=400, Evaluation Protocol=Linear2021.10 | 85.36 | — | — | — | — | — | — | — | — | — | — | — | — | |
| HolisticPUSetting=2, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 85.3 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BPNetwork=AlexNet, Setting=Muon + BN2026.06 | 85.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| ScatSimCLREvaluation protocol=Linear evaluation on frozen representations2023.03 | 85.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BPNetwork=AlexNet, Setting=Muon2026.06 | 85.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR†+IP-IRMTraining Epochs=1000, Evaluation Protocol=k-NN, MixUp=true2021.10 | 85.08 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineTraining Augmentation=AugMix2023.10 | 84.99 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR + DebiasedMethod Variant=Debiased2021.10 | 84.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| UnMixMatchLabeled Samples=10002023.06 | 84.73 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEConv-2Training Augmentation=None2023.10 | 84.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEConv-2Training Augmentation=Color Jitter2023.10 | 84.46 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLR+IP-IRMTraining Epochs=400, Evaluation Protocol=Linear2021.10 | 84.44 | — | — | — | — | — | — | — | — | — | — | — | — | |
| HCL+IP-IRMTraining Epochs=400, Evaluation Protocol=k-NN2021.10 | 84.29 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SimCLRTraining Epochs=1000, Evaluation Protocol=Linear2021.10 | 84.24 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEConvTraining Augmentation=None2023.10 | 84.24 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CEConvTraining Augmentation=Color Jitter2023.10 | 84.16 | — | — | — | — | — | — | — | — | — | — | — | — | |
| P³MIX-CSetting=2, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 84.1 | — | — | — | — | — | — | — | — | — | — | — | — | |
| α-CL-directBackbone=ResNet18, Epochs=500, Batchsize=128, Learning rate=0.01, Evaluation protocol=linear evaluation2022.01 | 84.05 | — | — | — | — | — | — | — | — | — | — | — | — | |
| alpha-CL-directBackbone=ResNet18, Epochs=500, Batchsize=1282022.01 | 84.05 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineTraining Augmentation=Color Jitter2023.10 | 83.91 | — | — | — | — | — | — | — | — | — | — | — | — | |
| BaselineTraining Augmentation=None2023.10 | 83.8 | — | — | — | — | — | — | — | — | — | — | — | — | |
| InfoNCE (Lnce)Backbone=ResNet18, Epochs=500, Batchsize=128, Learning rate=0.01, Evaluation protocol=linear evaluation2022.01 | 83.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| L_nceBackbone=ResNet18, Epochs=500, Batchsize=1282022.01 | 83.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| HolisticPUSetting=1, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 83.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| P³MIX-ESetting=2, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 83.7 | — | — | — | — | — | — | — | — | — | — | — | — | |
| PULNSSetting=2, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 83.6 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CLIPzero-shot=true2022.05 | 83.5 | — | — | — | — | — | — | — | — | — | — | — | — | |
| α-CL-directBackbone=ResNet50, Epochs=100, Batchsize=128, Learning rate=0.001, Evaluation protocol=linear evaluation2022.01 | 83.2 | — | — | — | — | — | — | — | — | — | — | — | — | |
| α-CL-directBackbone=ResNet18, Epochs=300, Batchsize=128, Learning rate=0.01, Evaluation protocol=linear evaluation2022.01 | 82.95 | — | — | — | — | — | — | — | — | — | — | — | — | |
| VPUSetting=2, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 82.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| Dist-PUSetting=2, Backbone=7-Layer CNN, Input size=3×96×962023.10 | 82.9 | — | — | — | — | — | — | — | — | — | — | — | — | |
| FlexMatchLabeled Samples=10002023.06 | 82.67 | — | — | — | — | — | — | — | — | — | — | — | — | |
| DCLTraining Epochs=400, Evaluation Protocol=Linear2021.10 | 82.56 | — | — | — | — | — | — | — | — | — | — | — | — | |
| SupervisedBackProp=-, EMA=-, Evaluation Protocol=Supervised2021.07 | 82.55 | — | — | — | — | — | — | — | — | — | — | — | — | |
| UDALabeled Samples=10002023.06 | 82.54 | — | — | — | — | — | — | — | — | — | — | — | — | |
| InfoNCE (Lnce)Backbone=ResNet18, Epochs=300, Batchsize=128, Learning rate=0.01, Evaluation protocol=linear evaluation2022.01 | 82.49 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CoMatchLabeled Samples=10002023.06 | 82 | — | — | — | — | — | — | — | — | — | — | — | — | |
| CCSSLLabeled Samples=10002023.06 | 82 | — | — | — | — | — | — | — | — | — | — | — | — |