AdaSCALE: Adaptive Scaling for OOD Detection
Sudarshan Regmi · 2025
Read AdaSCALE: Adaptive Scaling for OOD Detection by Sudarshan Regmi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-08023.txt.
Description
Sudarshan Regmi improves the test that tells a vision model it is looking at something it was never trained on, in a 2025 paper. Out-of-distribution detection matters for safe deployment, and current methods reshape a network's activations with one fixed percentile threshold for every input. Regmi noticed something about foreign inputs under slight disturbance. Their largest activations shift much more than familiar inputs' do.
His method, AdaSCALE, turns that observation into an estimate of how foreign each sample looks, then sets the scaling threshold per sample. Likely familiar inputs get strong scaling; likely foreign ones get weak scaling. The resulting scores separated the two classes far better, beating the strongest rival by wide margins on a standard benchmark across eight architectures.
Catalog record
| CREATOR | Sudarshan Regmi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 90 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-08023