Out-of-distribution (OOD) detection — approaches.
hardAnswer
- Softmax confidence is unreliable (over-confident on OOD).
- Better: (1) MSP with temperature scaling, (2) energy score (LeCun et al.), (3) ODIN (temperature + input perturbation), (4) Mahalanobis distance in feature space, (5) deep generative likelihood (with Nalisnick's caveat), (6) contrastive OOD detectors.
- Modern default: energy score or Mahalanobis on penultimate features.
- Critical in safety-critical deployment (medical, autonomous).
Check yourself — multiple choice
- Softmax alone
- MSP + temperature / energy score / ODIN / Mahalanobis on features / contrastive detectors; softmax overconfident on OOD — modern default: energy or Mahalanobis
- Random
- Not real
OOD: energy / Mahalanobis / ODIN / contrastive; softmax overconfident.
#anomaly-detection#deep-learning
Practise Unsupervised Learning
214 interview questions in this topic.