Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection

Shreen Gul, Mohamed Elmahallawy, Ardhendu Tripathy, Sanjay Madria · 2026

Read Prototype Fusion: A Training-Free Multi-Layer Approach to OOD Detection by Shreen Gul, Mohamed Elmahallawy, Ardhendu Tripathy, Sanjay Madria (2026) 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-2603-23677.txt.

Description

Gul and three co-authors detect out-of-distribution images without any training. Most detectors read only the last layer of a network before its output. The team show mid layers carry just as much signal for spotting odd inputs. Their method pools features from block after block and averages them per class into prototypes. A test image scores by its cosine closeness to the nearest prototype. Odd images sit far from all of them. Across standard benchmarks, the fusion lifts detection and cuts false alarms.

Catalog record

CREATORShreen Gul, Mohamed Elmahallawy, Ardhendu Tripathy, Sanjay Madria
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33314
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2603.23677

Browse the full library · About the Tower of Babel