Out-of-distribution Detection on SVHN (test)
0.99AUROCOOD EBM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OOD EBMsupervised=true, fine-tuned=true2022.10 | 0.99 | — | — | |
| O-Langevin2022.10 | 0.9868 | — | — | |
| SADA-JEMScore Function=log p_theta(x), Training Dataset=CIFAR10, K=202022.09 | 0.98 | — | — | |
| Tempered SGLD2022.10 | 0.9763 | — | — | |
| HDGETraining Protocol=Supervised2020.10 | 0.96 | — | — | |
| HDGEsupervised=true2022.10 | 0.96 | — | — | |
| SADA-JEMScore Function=max_y p_theta(y|x), Training Dataset=CIFAR10, K=202022.09 | 0.96 | — | — | |
| Supervisedsupervision level=full2021.05 | 0.956 | — | — | |
| SADA-JEMScore Function=log p_theta(x), Training Dataset=CIFAR10, K=102022.09 | 0.95 | — | — | |
| O-SVGD2022.10 | 0.9469 | — | — | |
| JEM++Score Function=max_y p_theta(y|x), Training Dataset=CIFAR10, M=202022.09 | 0.94 | — | — | |
| SVGD2022.10 | 0.9355 | — | — | |
| WideResNetScore Function=max_y p_theta(y|x), Training Dataset=CIFAR102022.09 | 0.93 | — | — | |
| SADA-JEMScore Function=max_y p_theta(y|x), Training Dataset=CIFAR10, K=102022.09 | 0.93 | — | — | |
| OpenMatchlabeled samples per class=100, supervision level=partial2021.05 | 0.93 | — | — | |
| Hat EBMsupervised=false2022.10 | 0.92 | — | — | |
| SADA-JEMScore Function=max_y p_theta(y|x), Training Dataset=CIFAR10, K=52022.09 | 0.92 | — | — | |
| Improved CD EBMsupervised=false2022.10 | 0.91 | — | — | |
| OOD EBMsupervised=true2022.10 | 0.91 | — | — | |
| WideResNetScore Function=log p_theta(x), Training Dataset=CIFAR102022.09 | 0.91 | — | — | |
| ImCDScore Function=log p_theta(x), Training Dataset=CIFAR102022.09 | 0.91 | — | — | |
| SADA-JEMScore Function=log p_theta(x), Training Dataset=CIFAR10, K=52022.09 | 0.91 | — | — | |
| SGLD2022.10 | 0.8941 | — | — | |
| Deterministic controlScore=ODIN, Seeds=5-seed hard-delivery2026.05 | 0.8921 | — | — | |
| JEMScore Function=max_y p_theta(y|x), Training Dataset=CIFAR10, K=202022.09 | 0.89 | — | — | |
| Deterministic controlScore=Energy, Seeds=5-seed hard-delivery2026.05 | 0.889 | — | — | |
| MultipassScore=ODIN, Seeds=5-seed hard-delivery2026.05 | 0.8871 | — | — | |
| MultipassScore=Energy, Seeds=5-seed hard-delivery2026.05 | 0.8858 | — | — | |
| SLSScore=ODIN, Seeds=5-seed hard-delivery2026.05 | 0.8812 | — | — | |
| Deterministic controlScore=Pred. entropy, Seeds=5-seed hard-delivery2026.05 | 0.8793 | — | — | |
| SLSScore=Energy, Seeds=5-seed hard-delivery2026.05 | 0.8788 | — | — | |
| MTClabeled samples per class=100, supervision level=partial2021.05 | 0.876 | — | — | |
| MultipassScore=Pred. entropy, Seeds=5-seed hard-delivery2026.05 | 0.8714 | — | — | |
| SLSScore=Pred. entropy, Seeds=5-seed hard-delivery2026.05 | 0.866 | — | — | |
| Soft labelsScore=ODIN, Seeds=10-seed2026.05 | 0.865 | — | — | |
| Deterministic controlScore=MSP, Seeds=5-seed hard-delivery2026.05 | 0.8643 | — | — | |
| Soft labelsScore=Energy, Seeds=10-seed2026.05 | 0.8634 | — | — | |
| MultipassScore=MSP, Seeds=5-seed hard-delivery2026.05 | 0.8553 | — | — | |
| Soft labelsScore=Pred. entropy, Seeds=10-seed2026.05 | 0.8527 | — | — | |
| SLSScore=MSP, Seeds=5-seed hard-delivery2026.05 | 0.852 | — | — | |
| JEM++Score Function=log p_theta(x), Training Dataset=CIFAR10, M=202022.09 | 0.85 | — | — | |
| Deterministic controlScore=Logit margin, Seeds=5-seed hard-delivery2026.05 | 0.8419 | — | — | |
| Soft labelsScore=MSP, Seeds=10-seed2026.05 | 0.8364 | — | — | |
| Labeled Onlylabeled samples per class=100, supervision level=partial2021.05 | 0.836 | — | — | |
| SLSScore=Logit margin, Seeds=5-seed hard-delivery2026.05 | 0.8326 | — | — | |
| MultipassScore=Logit margin, Seeds=5-seed hard-delivery2026.05 | 0.8323 | — | — | |
| VAEBMTraining Protocol=Unsupervised2020.10 | 0.83 | — | — | |
| VERAIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.83 | — | — | |
| VAEBMsupervised=false2022.10 | 0.83 | — | — | |
| VERAScore Function=log p_theta(x), Training Dataset=CIFAR102022.09 | 0.83 | — | — | |
| Soft labelsScore=Logit margin, Seeds=10-seed2026.05 | 0.8112 | — | — | |
| FixMatchlabeled samples per class=100, supervision level=partial2021.05 | 0.799 | — | — | |
| Soft labelsScore=KNN, Seeds=10-seed2026.05 | 0.7974 | — | — | |
| MultipassScore=KNN, Seeds=5-seed hard-delivery2026.05 | 0.7836 | — | — | |
| Deterministic controlScore=KNN, Seeds=5-seed hard-delivery2026.05 | 0.7788 | — | — | |
| SLSScore=KNN, Seeds=5-seed hard-delivery2026.05 | 0.7741 | — | — | |
| BiDVLIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.76 | — | — | |
| Divergence TriangleTraining Protocol=Unsupervised2020.10 | 0.68 | — | — | |
| Divergence triangleIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.68 | — | — | |
| JEMTraining Protocol=Supervised2020.10 | 0.67 | — | — | |
| JEMIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.67 | — | — | |
| JEMsupervised=true2022.10 | 0.67 | — | — | |
| JEMScore Function=log p_theta(x), Training Dataset=CIFAR10, K=202022.09 | 0.67 | — | — | |
| IGEBMTraining Protocol=Unsupervised2020.10 | 0.63 | — | — | |
| IGEBMIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.63 | — | — | |
| IGEBMScore Function=log p_theta(x), Training Dataset=CIFAR102022.09 | 0.63 | — | — | |
| IGEBMsupervised=false2022.10 | 0.43 | — | — | |
| IGEBMScore Function=max_y p_theta(y|x), Training Dataset=CIFAR102022.09 | 0.43 | — | — | |
| NVAETraining Protocol=Unsupervised2020.10 | 0.42 | — | — | |
| SVAEIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.42 | — | — | |
| GlowIn-distribution dataset=CIFAR-10, OOD score=Negative free energy2022.03 | 0.24 | — | — | |
| GlowTraining Protocol=Unsupervised2020.10 | 0.05 | — | — | |
| Input ComplexityModel class=Diffusion2023.12 | — | 0.973 | 0.976 | |
| Input LikelihoodModel class=Diffusion2023.12 | — | 0.974 | 0.97 | |
| Likelihood RegretModel class=Diffusion2023.12 | — | 0.805 | 0.821 | |
| LMDModel class=Diffusion2023.12 | — | 0.914 | 0.876 | |
| LMDModel class=Consistency2023.12 | — | 0.832 | 0.792 | |
| MSMAModel class=Diffusion2023.12 | — | 0.976 | 0.979 | |
| MSMAModel class=Consistency2023.12 | — | 0.985 | 0.981 | |
| Projection RegretModel class=Consistency2023.12 | — | 0.995 | 0.993 |