Research map / Label-efficient and robust learning / OOD and anomaly detection
Towards Maximizing the Representation Gap between In-Domain & Out-of-Distribution Examples
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use negative λout to train OOD examples toward sharp multi-modal Dirichlet outputs | DPN · C10, C100, and TIM OOD detection | AUROC (MI): 97.8 ±0.4 -> 99.0 ±0.1 (C10/Tiny); 92.2 ±0.1 -> 94.5 ±0.1 (C100/Tiny); 99.7 ±0.0 -> 99.9 ±0.0 (TIM/CIFAR-10) | Helped |
| Use negative λout to train OOD examples toward sharp multi-modal Dirichlet outputs | DPN · C10 misclassification detection | AUROC (MI): 90.3 ±0.1 -> 89.9 ±0.0 | Hurt |
| Use negative λout to train OOD examples toward sharp multi-modal Dirichlet outputs | DPN · C100 and TIM misclassification detection | AUROC (MI): 81.2 ±0.0 -> 82.3 ±0.0 (C100); 78.3 ±0.4 -> 79.1 ±0.5 (TIM) | Helped |
| Use negative λout, which makes OOD Dirichlet outputs sharp and multi-modal | DPN · C10, C100, and TIM OOD detection | AUROC (D.Ent): 97.7 ±0.4 -> 6.0 ±0.3 (C10/Tiny); 92.3 ±0.1 -> 38.1 ±0.5 (C100/Tiny); 99.6 ±0.0 -> 3.5 ±0.1 (TIM/CIFAR-10) | Hurt |
| Set λin to 0 while keeping λout at -0.5 (ablation) | DPN− · C100 OOD detection | AUROC (MI): 94.5 ±0.1 -> 91.1 ±0.1 (Tiny); 96.8 ±0.1 -> 95.0 ±0.0 (STL-10); 96.5 ±0.1 -> 93.8 ±0.1 (LSUN) | Hurt |
| Set λout to 0 while keeping λin at 0.5 (ablation) | DPN− · C100 OOD detection | AUROC (MI): 94.5 ±0.1 -> 93.6 ±0.1 (Tiny); 96.8 ±0.1 -> 96.2 ±0.3 (STL-10); 96.5 ±0.1 -> 95.9 ±0.1 (LSUN) | Hurt |
| Set the OOD target concentration parameters to 0.1 for every class in the RKL loss (ablation) | DPNrev · C100 OOD detection | AUROC (MI): 83.8 ±0.1 -> 80.4 ±0.2 (Tiny); 89.3 ±0.1 -> 84.3 ±0.1 (STL-10); 89.3 ±0.1 -> 82.2 ±0.0 (LSUN) | Hurt |
| Use a binary in-domain-versus-OOD classifier instead of DPN− | OOD detection · C10 / STL-10 | AUROC: 85.3 ±0.5 -> 93.7 ±0.9 | Helped |
| Use a binary in-domain-versus-OOD classifier instead of DPN− | OOD detection · C100 / Tiny, STL-10, LSUN, Places365, and Textures | AUROC: 94.5 ±0.1 -> 84.8 ±1.2 (Tiny); 96.8 ±0.1 -> 90.7 ±0.9 (STL-10); 96.5 ±0.1 -> 91.2 ±1.2 (LSUN); 94.5 ±0.1 -> 88.2 ±1.3 (P | Hurt |
Closest papers by meaning
- Post-hoc Uncertainty Learning Using a Dirichlet Meta-Model (AAAI 2023 · similarity 0.88)
- Adaptive Uncertainty Estimation via High-Dimensional Testing on Latent Representations (NeurIPS 2023 · similarity 0.87)
- Out of Distribution Data Detection Using Dropout Bayesian Neural Networks (AAAI 2022 · similarity 0.87)
- Predictive Uncertainty Estimation via Prior Networks (NeurIPS 2018 · similarity 0.86)
- The Best of Both Worlds: On the Dilemma of Out-of-distribution Detection (NeurIPS 2024 · similarity 0.85)
- Weakly-Supervised Residual Evidential Learning for Multi-Instance Uncertainty Estimation (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.