Image Classification on CIFAR10 (test)
98.1AccuracyLinear Probe
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Linear ProbeInterpretability=false, Evaluation Setting=Fully-supervised, Backbone=CLIP ViT-L/142026.01 | 98.1 | — | |
| PCBM-ReDInterpretability=true, Evaluation Setting=Fully-supervised, Backbone=CLIP ViT-L/142026.01 | 98.05 | — | |
| V2C-CBMInterpretability=true, Evaluation Setting=Fully-supervised, Backbone=CLIP ViT-L/142026.01 | 98.03 | — | |
| Res-CBMInterpretability=true, Evaluation Setting=Fully-supervised, Backbone=CLIP ViT-L/142026.01 | 97.77 | — | |
| LaBoInterpretability=true, Evaluation Setting=Fully-supervised, Backbone=CLIP ViT-L/142026.01 | 97.75 | — | |
| SSAM-DBackbone=WideResNet-28-10, Sparsity=80%, FLOPs=1.44x2022.10 | 97.72 | — | |
| SSAM-FBackbone=WideResNet-28-10, Sparsity=50%, FLOPs=1.65x2022.10 | 97.71 | — | |
| SSAM-DBackbone=WideResNet-28-10, Sparsity=50%, FLOPs=1.65x2022.10 | 97.7 | — | |
| SSAM-FBackbone=WideResNet-28-10, Sparsity=80%, FLOPs=1.44x2022.10 | 97.67 | — | |
| SSAM-FBackbone=WideResNet-28-10, Sparsity=99%, FLOPs=1.30x2022.10 | 97.59 | — | |
| SSAM-DBackbone=WideResNet-28-10, Sparsity=90%, FLOPs=1.36x2022.10 | 97.53 | — | |
| SSAM-DBackbone=WideResNet-28-10, Sparsity=95%, FLOPs=1.33x2022.10 | 97.52 | — | |
| SAMBackbone=WideResNet-28-10, Sparsity=0%, FLOPs=2x2022.10 | 97.48 | — | |
| SSAM-FBackbone=WideResNet-28-10, Sparsity=90%, FLOPs=1.36x2022.10 | 97.47 | — | |
| SSAM-FBackbone=WideResNet-28-10, Sparsity=95%, FLOPs=1.33x2022.10 | 97.42 | — | |
| SSAM-FBackbone=WideResNet-28-10, Sparsity=98%, FLOPs=1.31x2022.10 | 97.32 | — | |
| SSAM-DBackbone=WideResNet-28-10, Sparsity=98%, FLOPs=1.31x2022.10 | 97.3 | — | |
| SSAM-DBackbone=WideResNet-28-10, Sparsity=99%, FLOPs=1.30x2022.10 | 97.27 | — | |
| SGDBackbone=WideResNet-28-10, Sparsity=N/A, FLOPs=1x2022.10 | 97.11 | — | |
| Supervised CLIPTraining Protocol=Supervised2021.06 | 96.7 | — | |
| CuPLInterpretability=false, Evaluation Setting=Zero-shot, Backbone=CLIP ViT-L/142026.01 | 95.82 | — | |
| PCBM-ReD + CuPLInterpretability=true, Evaluation Setting=Zero-shot, Backbone=CLIP ViT-L/142026.01 | 95.8 | — | |
| CLIPInterpretability=false, Evaluation Setting=Zero-shot, Backbone=CLIP ViT-L/142026.01 | 95.57 | — | |
| PCBM-ReDInterpretability=true, Evaluation Setting=Zero-shot, Backbone=CLIP ViT-L/142026.01 | 95.56 | — | |
| CLIP + EB βTraining Protocol=Zero-shot SSL compressor2021.06 | 95.2 | 7 | |
| MixFixBackbone=PreResNet-18, Zero-shot=false2024.08 | 95.15 | — | |
| SGDmBackbone=ResNet-34, LR Schedule=Decayed2022.02 | 95.08 | — | |
| APOBackbone=ResNet-34, Base Optimizer=SGDm2022.02 | 94.47 | — | |
| APOBackbone=ResNet-34, Base Optimizer=SGD2022.02 | 94.27 | — | |
| AdamBackbone=ResNet-34, LR Schedule=Decayed2022.02 | 94.12 | — | |
| SSAM-FBackbone=VGG11-BN, Sparsity=50%, Perturbation magnitude ρ=0.052022.10 | 94.03 | — | |
| RMSpropBackbone=ResNet-34, LR Schedule=Decayed2022.02 | 93.97 | — | |
| APOBackbone=ResNet-34, Base Optimizer=RMSprop2022.02 | 93.97 | — | |
| SSAM-DBackbone=VGG11-BN, Sparsity=80%, Perturbation magnitude ρ=0.052022.10 | 93.95 | — | |
| SAMBackbone=VGG11-BN, Sparsity=0%, Perturbation magnitude ρ=0.052022.10 | 93.87 | — | |
| SSAM-DBackbone=VGG11-BN, Sparsity=90%, Perturbation magnitude ρ=0.052022.10 | 93.85 | — | |
| SSAM-FBackbone=VGG11-BN, Sparsity=80%, Perturbation magnitude ρ=0.052022.10 | 93.83 | — | |
| APOBackbone=ResNet-34, Base Optimizer=Adam2022.02 | 93.8 | — | |
| SSAM-DBackbone=VGG11-BN, Sparsity=50%, Perturbation magnitude ρ=0.052022.10 | 93.79 | — | |
| SSAM-FBackbone=VGG11-BN, Sparsity=95%, Perturbation magnitude ρ=0.052022.10 | 93.77 | — | |
| SSAM-FBackbone=VGG11-BN, Sparsity=90%, Perturbation magnitude ρ=0.052022.10 | 93.76 | — | |
| SGDBackbone=ResNet-34, LR Schedule=Decayed2022.02 | 93.54 | — | |
| SSAM-FBackbone=VGG11-BN, Sparsity=98%, Perturbation magnitude ρ=0.052022.10 | 93.54 | — | |
| SSAM-DBackbone=VGG11-BN, Sparsity=98%, Perturbation magnitude ρ=0.052022.10 | 93.54 | — | |
| SSAM-DBackbone=VGG11-BN, Sparsity=95%, Perturbation magnitude ρ=0.052022.10 | 93.48 | — | |
| SSAM-FBackbone=VGG11-BN, Sparsity=99%, Perturbation magnitude ρ=0.052022.10 | 93.47 | — | |
| SGDBackbone=VGG11-BN, Sparsity=N/A, Perturbation magnitude ρ=0.052022.10 | 93.42 | — | |
| SSAM-DBackbone=VGG11-BN, Sparsity=99%, Perturbation magnitude ρ=0.052022.10 | 93.33 | — | |
| AdamBackbone=ResNet-34, LR Schedule=Fixed2022.02 | 93.23 | — | |
| SGDBackbone=ResNet-34, LR Schedule=Fixed2022.02 | 93 | — | |
| SGDmBackbone=ResNet-34, LR Schedule=Fixed2022.02 | 92.99 | — | |
| RMSpropBackbone=ResNet-34, LR Schedule=Fixed2022.02 | 92.87 | — | |
| SOTA [22]Backbone=PreResNet-18, Zero-shot=false2024.08 | 92.68 | — | |
| CLIPBackbone=ViT-B/32, Zero-shot=true2024.08 | 89.97 | — | |
| FedEM2021.08 | 84.3 | — | |
| FedAvg+2021.08 | 82.3 | — | |
| FedEMclient sampling=20%, communication rounds=12002021.08 | 82.1 | — | |
| pFedMe2021.08 | 81.7 | — | |
| Clustered FL2021.08 | 78.6 | — | |
| FedAvg2021.08 | 78.2 | — | |
| APFLclient sampling=20%, communication rounds=12002021.08 | 78.2 | — | |
| pFedMeclient sampling=20%, communication rounds=12002021.08 | 77.8 | — | |
| FedAvg+client sampling=20%, communication rounds=12002021.08 | 77.7 | — | |
| D-FedEMtopology=Binomial Erdős-Rényi graph (p=0.5), communication=fully decentralized2021.08 | 77 | — | |
| FedAvgclient sampling=20%, communication rounds=12002021.08 | 73.1 | — | |
| DP-RandPEpsilon (ε)=3, Evaluation Protocol=Linear probing2023.06 | 71.08 | — | |
| Local2021.08 | 70.2 | — | |
| DP-RandPEpsilon (ε)=2, Evaluation Protocol=Linear probing2023.06 | 69.92 | — | |
| Tramèr and Boneh (ScatterNet)Epsilon (ε)=3, Evaluation Protocol=CNN on top of extracted features2023.06 | 69.3 | — | |
| DP-RandPEpsilon (ε)=1, Evaluation Protocol=Linear probing2023.06 | 67.78 | — | |
| Tramèr and Boneh (ScatterNet)Epsilon (ε)=2, Evaluation Protocol=CNN on top of extracted features2023.06 | 67.2 | — | |
| DP-RandPEpsilon (ε)=0.5, Evaluation Protocol=Linear probing2023.06 | 65.1 | — | |
| DP-RandPEpsilon (ε)=0.2, Evaluation Protocol=Linear probing2023.06 | 60.89 | — | |
| Tramèr and Boneh (ScatterNet)Epsilon (ε)=1, Evaluation Protocol=CNN on top of extracted features2023.06 | 60.3 | — | |
| DP-RandPEpsilon (ε)=0.1, Evaluation Protocol=Linear probing2023.06 | 57.1 | — | |
| DP-RandPEpsilon (ε)=0.03, Evaluation Protocol=Linear probing2023.06 | 40.64 | — |