ODIN: temperature and input-perturbation OOD detection

What changed

Combines high-temperature softmax scores with small input-gradient perturbations to separate known and anomalous images while leaving the underlying classifier weights unchanged.

Scientific scope

Combines high-temperature softmax scores with small input-gradient perturbations to separate known and anomalous images while leaving the underlying classifier weights unchanged. Requires selected temperature, perturbation and threshold; original experiments tune on held-out examples from each OOD distribution. Theory uses additional score/distribution conditions, not arbitrary future-anomaly guarantees. Reported FPR uses in-distribution positives, unlike some later conventions.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Requires selected temperature, perturbation and threshold; original experiments tune on held-out examples from each OOD distribution. Theory uses additional score/distribution conditions, not arbitrary future-anomaly guarantees. Reported FPR uses in-distribution positives, unlike some later conventions.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Principled Detection of Out-of-Distribution Examples in Neural Networks

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction