OOD Detection on Dirty-MNIST vs Fashion MNIST (test)
0.271NLLENSEMBLE
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ENSEMBLEBackbone=LeNet, Runs=52024.10 | 0.271 | 83.915 | 1.306 | 86.095 | 96.065 | |
| SINGLEBackbone=LeNet, Runs=52024.10 | 0.272 | 82.299 | 2.908 | 65.935 | 50 | |
| ENSEMBLE+OOD HE-CKABackbone=LeNet, Runs=5, OOD Tuning=OOD Likelihood Minimization2024.10 | 0.277 | 84.09 | 1.712 | 99.996 | 99.742 | |
| DDUBackbone=LeNet, Runs=5, OOD Metric=Feature space density (for PE)2024.10 | 0.278 | 82.177 | 3.952 | 94.168 | — | |
| HYPERNETBackbone=LeNet, Runs=52024.10 | 0.278 | 81.157 | 4.253 | 46.393 | 64.856 | |
| SVGD+HE-CKABackbone=LeNet, Runs=5, Kernel=Cosine similarity feature kernel2024.10 | 0.298 | 83.879 | 2.846 | 94.38 | 99.213 | |
| SVGD+RBFBackbone=LeNet, Runs=5, Kernel=RBF2024.10 | 0.304 | 83.56 | 3.178 | 91.003 | 98.083 | |
| ENSEMBLE+HE-CKABackbone=LeNet, Runs=52024.10 | 0.306 | 83.684 | 3.174 | 94.656 | 98.866 | |
| HYPERNET+OOD HE-CKABackbone=LeNet, Runs=5, OOD Tuning=OOD Likelihood Minimization2024.10 | 0.325 | 82.398 | 3.058 | 98.073 | 77.548 | |
| SVGD+CKApwBackbone=LeNet, Runs=5, Kernel=Cosine similarity feature kernel2024.10 | 0.377 | 82.351 | 7.359 | 89.195 | 99.207 |