Out-of-distribution Detection on CIFAR-100 (ID) vs TinyImageNet (OOD)
86.24OOD AUROCLate-phase BatchNorm (non-averaged)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Late-phase BatchNorm (non-averaged)Backbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=10, Weight Averaging=false2020.07 | 86.24 | — | — | |
| Deep Ensembles (Late-phase BatchNorm)Backbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=102020.07 | 85.75 | — | — | |
| Late-phase BatchNormBackbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=1, Weight Averaging=true2020.07 | 83.6 | — | — | |
| Deep EnsemblesBackbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=102020.07 | 83.12 | — | — | |
| BatchEnsembleBackbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=102020.07 | 82.85 | — | — | |
| SWAGBackbone=WRN-28-10, Optimizer=SWA, Ensemble Size (K)=102020.07 | 82.83 | — | — | |
| MC-DropoutBackbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=102020.07 | 82.25 | — | — | |
| Late-phase HypernetworkBackbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=12020.07 | 82.09 | — | — | |
| Dropout (Mean)Backbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=12020.07 | 80.22 | — | — | |
| BaseBackbone=WRN-28-10, Optimizer=SGD, Ensemble Size (K)=12020.07 | 80.15 | — | — | |
| Deep Abstaining ClassifierBackbone=Wide ResNet 28x102021.05 | — | 18.68 | 94.88 | |
| Deep Abstaining ClassifierBackbone=ResNet 342021.05 | — | 12.1 | 97.14 | |
| Deep Mahalanobis DetectorBackbone=ResNet 342021.05 | — | 29.7 | 87.9 | |
| Ensemble of Leave-out ClassifiersBackbone=Wide ResNet 28x102021.05 | — | 24.53 | 95.18 | |
| ODINBackbone=Wide ResNet 28x102021.05 | — | 55.9 | 84 | |
| OpenMaxBackbone=ResNet 342021.05 | — | 32.67 | 81.22 |