Image Classification on CIFAR-10
99.61AccuracyEfficient Adaptive Ensembling (Full method)
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Efficient Adaptive Ensembling (Full method)2022.06 | 99.61 | — | — | — | — | — | — | — | — | — | — | |
| DINOv2Architecture=ViT-g/14, Pre-training Data=LVD, Resolution=518x518, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 99.6 | — | — | — | — | — | — | — | — | — | — | |
| Transformer architecture2024.06 | 99.5 | — | — | — | — | — | — | — | — | — | — | |
| CaiT-M-36 ↑ 224Resolution=224, Crop-ratio=0.875, Training=Longer schedules2021.03 | 99.4 | — | — | — | — | — | — | — | — | — | — | |
| CaiT-M-36 224Resolution=224, Crop-ratio=0.8752021.03 | 99.3 | — | — | — | — | — | — | — | — | — | — | |
| CaiT-S-36 224Resolution=224, Crop-ratio=0.8752021.03 | 99.2 | — | — | — | — | — | — | — | — | — | — | |
| CaiT-S-36 ↑ 224Resolution=224, Crop-ratio=0.875, Training=Longer schedules2021.03 | 99.2 | — | — | — | — | — | — | — | — | — | — | |
| iBOTArchitecture=ViT-L/16, Pre-training Data=IN-21k, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 99.2 | — | — | — | — | — | — | — | — | — | — | |
| DeiT-B 224Resolution=224, Crop-ratio=0.8752021.03 | 99.1 | — | — | — | — | — | — | — | — | — | — | |
| GPipeNumber of Parameters=556M, Pre-training=ImageNet, Transfer Learning=true2019.05 | 99 | — | — | — | — | — | — | — | — | — | — | |
| EfficientNet-B7Number of Parameters=64M, Pre-training=ImageNet, Transfer Learning=true2019.05 | 98.9 | — | — | — | — | — | — | — | — | — | — | |
| EfficientNet-B7Resolution=224, Crop-ratio=0.8752021.03 | 98.9 | — | — | — | — | — | — | — | — | — | — | |
| AIM-7B†Architecture=ViT-7B/14, Pre-training Data=DFN-2B+, Feature Extraction Layer=20th, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 98.9 | — | — | — | — | — | — | — | — | — | — | |
| FMixModel=Pyramid2020.02 | 98.64 | — | — | — | — | — | — | — | — | — | — | |
| SimCLRProtocol=Fine-tuned, Backbone=ResNet-50 (4x), Pre-training=ImageNet2020.02 | 98.6 | — | — | — | — | — | — | — | — | — | — | |
| Dom-Ad (Am-B)Backbone Architecture=AmoebaNet-B2020.09 | 98.6 | — | — | — | — | — | — | — | — | — | — | |
| Dom-AdBackbone=AmoebaNet-B2020.09 | 98.6 | — | — | — | — | — | — | — | — | — | — | |
| AIM-7BArchitecture=ViT-7B/14, Pre-training Data=DFN-2B+, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 98.6 | — | — | — | — | — | — | — | — | — | — | |
| AIM-3BArchitecture=ViT-3B/14, Pre-training Data=DFN-2B+, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 98.4 | — | — | — | — | — | — | — | — | — | — | |
| BaselineModel=Pyramid2020.02 | 98.31 | — | — | — | — | — | — | — | — | — | — | |
| JFT - Adaptive TransferPre-training Source=JFT, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 98.3 | — | — | — | — | — | — | — | — | — | — | |
| SupervisedProtocol=Fine-tuned, Backbone=ResNet-50 (4x), Pre-training=ImageNet2020.02 | 98.3 | — | — | — | — | — | — | — | — | — | — | |
| Dom-Ad (In-v3)Backbone Architecture=Inception-v32020.09 | 98.3 | — | — | — | — | — | — | — | — | — | — | |
| Dom-AdBackbone=Inception-v32020.09 | 98.3 | — | — | — | — | — | — | — | — | — | — | |
| CutMixModel=Pyramid2020.02 | 98.24 | — | — | — | — | — | — | — | — | — | — | |
| AIM-1BArchitecture=ViT-1B/14, Pre-training Data=DFN-2B+, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 98.2 | — | — | — | — | — | — | — | — | — | — | |
| JFT - AnimalPre-training Source=JFT, Subset=Animal, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 98.1 | — | — | — | — | — | — | — | — | — | — | |
| EfficientNet-B0Number of Parameters=4M, Pre-training=ImageNet, Transfer Learning=true2019.05 | 98.1 | — | — | — | — | — | — | — | — | — | — | |
| ViT-B/16Resolution=224, Crop-ratio=0.8752021.03 | 98.1 | — | — | — | — | — | — | — | — | — | — | |
| NASNet-ANumber of Parameters=85M, Pre-training=ImageNet, Transfer Learning=true2019.05 | 98 | — | — | — | — | — | — | — | — | — | — | |
| MixUpModel=Pyramid2020.02 | 97.92 | — | — | — | — | — | — | — | — | — | — | |
| KNNSelection Method=KNN, Adaptation Strategy=Adapters, Training Epochs=4e, Backbone Architecture=Resnet-50-v22020.09 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| KNNSelection Method=KNN, Adaptation Strategy=Full, Training Epochs=2e, Backbone Architecture=Resnet-50-v22020.09 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| KLSelection Method=KL, Adaptation Strategy=Adapters, Training Epochs=4e, Backbone Architecture=Resnet-50-v22020.09 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| KLSelection Method=KL, Adaptation Strategy=Full, Training Epochs=2e, Backbone Architecture=Resnet-50-v22020.09 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| KLSelection Method=KL, Adaptation Strategy=Full, Training Epochs=4e, Backbone Architecture=Resnet-50-v22020.09 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| AdaptersBackbone=ResNet-50, Pre-training=JFT, Protocol=Adapters2020.09 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| ViT-L/16Resolution=224, Crop-ratio=0.8752021.03 | 97.9 | — | — | — | — | — | — | — | — | — | — | |
| AmoebaNet-B + cutoutParams(M)=34.9, Search Method=Evolution, Search Cost (GPU days)=31502020.06 | 97.87 | — | — | — | — | — | — | — | — | — | — | |
| BaselineBackbone Architecture=Resnet-50-v22020.09 | 97.8 | — | — | — | — | — | — | — | — | — | — | |
| KNNSelection Method=KNN, Adaptation Strategy=Full, Training Epochs=4e, Backbone Architecture=Resnet-50-v22020.09 | 97.8 | — | — | — | — | — | — | — | — | — | — | |
| BaselineBackbone=ResNet-502020.09 | 97.8 | — | — | — | — | — | — | — | — | — | — | |
| FullBackbone=ResNet-50, Pre-training=JFT, Protocol=Full2020.09 | 97.8 | — | — | — | — | — | — | — | — | — | — | |
| DINOArchitecture=ViT-B/8, Pre-training Data=IN-1k, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 97.8 | — | — | — | — | — | — | — | — | — | — | |
| ImageNet - Adaptive TransferPre-training Source=ImageNet, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 97.7 | — | — | — | — | — | — | — | — | — | — | |
| SimCLREvaluation protocol=Fine-tuned, Pre-training=ImageNet, Backbone=ResNet-502020.02 | 97.7 | — | — | — | — | — | — | — | — | — | — | |
| DARTS+ with cutoutParams(M)=3.7, Search Method=GB, Search Cost (GPU days)=0.42020.06 | 97.68 | — | — | — | — | — | — | — | — | — | — | |
| Entire JFT DatasetPre-training Source=JFT, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 97.6 | — | — | — | — | — | — | — | — | — | — | |
| Fine-Tuning DARTSBackbone=Ours, Augmentation=cutout, Params(M)=3.9, Search Method=GB, Search Cost (GPU days)=12020.06 | 97.52 | — | — | — | — | — | — | — | — | — | — | |
| SupervisedEvaluation protocol=Fine-tuned, Pre-training=ImageNet, Backbone=ResNet-502020.02 | 97.5 | — | — | — | — | — | — | — | — | — | — | |
| PDARTS + cutoutParams(M)=3.4, Search Method=GB, Search Cost (GPU days)=0.32020.06 | 97.5 | — | — | — | — | — | — | — | — | — | — | |
| P-DARTSParams (M)=3.4, FLOPs (M)=532, Cost (GPU Days)=0.32020.09 | 97.5 | — | — | — | — | — | — | — | — | — | — | |
| DARTS- (best)Params (M)=3.5, FLOPs (M)=568, Cost (GPU Days)=0.4, Search strategy (S0)=true2020.09 | 97.5 | — | — | — | — | — | — | — | — | — | — | |
| MAEArchitecture=ViT-2B/14, Pre-training Data=IG-3B, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 97.5 | — | — | — | — | — | — | — | — | — | — | |
| Efficient Adaptive Ensembling (Best adaptive ensemble)2022.06 | 97.49 | — | — | — | — | — | — | — | — | — | — | |
| PC-DARTS + cutoutParams(M)=3.6, Search Method=GB, Search Cost (GPU days)=0.12020.06 | 97.43 | — | — | — | — | — | — | — | — | — | — | |
| PC-DARTSParams (M)=3.6, FLOPs (M)=558, Cost (GPU Days)=0.12020.09 | 97.43 | — | — | — | — | — | — | — | — | — | — | |
| ImageNet - Entire DatasetPre-training Source=ImageNet, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 97.4 | — | — | — | — | — | — | — | — | — | — | |
| NASNetA + cutoutParams(M)=3.3, Search Method=RL, Search Cost (GPU days)=20002020.06 | 97.35 | — | — | — | — | — | — | — | — | — | — | |
| NASNet-AParams (M)=3.3, FLOPs (M)=608, Cost (GPU Days)=20002020.09 | 97.35 | — | — | — | — | — | — | — | — | — | — | |
| SETN + cutoutParams(M)=4.6, Search Method=GB, Search Cost (GPU days)=1.82020.06 | 97.31 | — | — | — | — | — | — | — | — | — | — | |
| FMixModel=Dense2020.02 | 97.3 | — | — | — | — | — | — | — | — | — | — | |
| EPNSelection Method=EPN, Adaptation Strategy=Adapters, Training Epochs=4e, Backbone Architecture=Resnet-50-v22020.09 | 97.3 | — | — | — | — | — | — | — | — | — | — | |
| AIM-0.6BArchitecture=ViT-H/14, Pre-training Data=DFN-2B+, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 97.2 | — | — | — | — | — | — | — | — | — | — | |
| ENASParams (M)=4.6, FLOPs (M)=626, Cost (GPU Days)=0.52020.09 | 97.11 | — | — | — | — | — | — | — | — | — | — | |
| MAEArchitecture=ViT-H/14, Pre-training Data=IN-1k, Evaluation Protocol=Attentive Probing, Backbone State=Frozen2024.01 | 97.1 | — | — | — | — | — | — | — | — | — | — | |
| GDASParams (M)=3.4, FLOPs (M)=519, Cost (GPU Days)=0.22020.09 | 97.07 | — | — | — | — | — | — | — | — | — | — | |
| MixUpModel=Dense2020.02 | 97.05 | — | — | — | — | — | — | — | — | — | — | |
| SNAS (mild) + cutoutParams(M)=2.9, Search Method=GB, Search Cost (GPU days)=1.52020.06 | 97.02 | — | — | — | — | — | — | — | — | — | — | |
| CutMixModel=Dense2020.02 | 96.96 | — | — | — | — | — | — | — | — | — | — | |
| JFT - BirdPre-training Source=JFT, Subset=Bird, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 96.9 | — | — | — | — | — | — | — | — | — | — | |
| EPNSelection Method=EPN, Adaptation Strategy=Full, Training Epochs=2e, Backbone Architecture=Resnet-50-v22020.09 | 96.9 | — | — | — | — | — | — | — | — | — | — | |
| NAONetParams(M)=10.6, Search Method=NAO, Search Cost (GPU days)=2002020.06 | 96.82 | — | — | — | — | — | — | — | — | — | — | |
| MixUpModel=WRN2020.02 | 96.6 | — | — | — | — | — | — | — | — | — | — | |
| Conditional Channel Gated NetworksEvaluation Protocol=Task-IL, Number of runs=52020.03 | 96.6 | — | — | — | — | — | — | — | — | — | — | |
| EPNSelection Method=EPN, Adaptation Strategy=Full, Training Epochs=4e, Backbone Architecture=Resnet-50-v22020.09 | 96.6 | — | — | — | — | — | — | — | — | — | — | |
| DenseNet-BCParams(M)=25.6, Search Method=Manual2020.06 | 96.54 | — | — | — | — | — | — | — | — | — | — | |
| CutMixModel=WRN2020.02 | 96.53 | — | — | — | — | — | — | — | — | — | — | |
| SANEEvaluation Protocol=linear evaluation, Augmentation=strong2021.06 | 96.5 | — | — | — | — | — | — | — | — | — | — | |
| CleanDefense=Cutmix2026.03 | 96.43 | — | — | — | — | — | — | — | — | — | — | |
| JFT - VehiclePre-training Source=JFT, Subset=Vehicle, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 96.4 | — | — | — | — | — | — | — | — | — | — | |
| JFT - FoodPre-training Source=JFT, Subset=Food, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 96.4 | — | — | — | — | — | — | — | — | — | — | |
| HATEvaluation Protocol=Task-IL, Number of runs=52020.03 | 96.4 | — | — | — | — | — | — | — | — | — | — | |
| FMixModel=WRN2020.02 | 96.38 | — | — | — | — | — | — | — | — | — | — | |
| BaselineModel=Dense2020.02 | 96.26 | — | — | — | — | — | — | — | — | — | — | |
| GDAS + cutoutParams(M)=2.5, Search Method=GB, Search Cost (GPU days)=0.172020.06 | 96.25 | — | — | — | — | — | — | — | — | — | — | |
| JFT - TransportPre-training Source=JFT, Subset=Transport, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 96.2 | — | — | — | — | — | — | — | — | — | — | |
| FMixModel=ResNet2020.02 | 96.14 | — | — | — | — | — | — | — | — | — | — | |
| SANEEvaluation Protocol=linear evaluation, Augmentation=weak2021.06 | 96.1 | — | — | — | — | — | — | — | — | — | — | |
| JFT - AircraftPre-training Source=JFT, Subset=Aircraft, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 96.1 | — | — | — | — | — | — | — | — | — | — | |
| DivideMixNoise Type=Symmetric, Noise Ratio=20%, Backbone=PreAct ResNet-182020.06 | 96.1 | — | — | — | — | — | — | — | — | — | — | |
| CutMixModel=ResNet2020.02 | 96 | — | — | — | — | — | — | — | — | — | — | |
| Random initProtocol=Fine-tuned, Backbone=ResNet-50 (4x), Pre-training=None2020.02 | 96 | — | — | — | — | — | — | — | — | — | — | |
| SMASHv2Params(M)=16, Search Method=GB, Search Cost (GPU days)=1.52020.06 | 95.97 | — | — | — | — | — | — | — | — | — | — | |
| i-Mix (+MoCo)Evaluation Protocol=linear evaluation, Augmentation=weak2021.06 | 95.9 | — | — | — | — | — | — | — | — | — | — | |
| Random initEvaluation protocol=Fine-tuned, Backbone=ResNet-502020.02 | 95.9 | — | — | — | — | — | — | — | — | — | — | |
| CleanDefense=Mixup2026.03 | 95.83 | — | — | — | — | — | — | — | — | — | — | |
| ELR+*Noise Type=Symmetric, Noise Ratio=20%, Backbone=PreAct ResNet-18, Evaluation Protocol=Highest validation accuracy during training2020.06 | 95.8 | — | — | — | — | — | — | — | — | — | — | |
| Random InitializationPre-training Source=None, Backbone=Inception v3, Evaluation Protocol=Fine-tuning2018.11 | 95.7 | — | — | — | — | — | — | — | — | — | — | |
| SupervisedProtocol=Linear evaluation, Backbone=ResNet-50 (4x), Pre-training=ImageNet2020.02 | 95.7 | — | — | — | — | — | — | — | — | — | — |