FEVER-OOD: Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection

Brian K.S. Isaac-Medina, Mauricio Che, Yona F.A. Gaus, Samet Akcay, Toby P. Brekcon · 2024

Read FEVER-OOD: Free Energy Vulnerability Elimination for Robust Out-of-Distribution Detection by Brian K.S. Isaac-Medina, Mauricio Che, Yona F.A. Gaus, Samet Akcay, Toby P. Brekcon (2024) 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-2412-01596.txt.

Description

Brian Isaac-Medina and co-authors fixed a blind spot in a way to spot inputs that fall outside a model's training data, in a 2024 paper. Vision models can be boldly wrong on such inputs. A score based on free energy had proved good at flagging them, but the team found a flaw in its math.

Two inputs can look far apart inside the model yet earn the same score, if their gap lies in a dead zone of the last layer. The fix shrinks the space used and pushes up the weakest response of that layer, which splits the scores apart. Added to a strong base model, the fix set a new best mark on a standard image test.

Catalog record

CREATORBrian K.S. Isaac-Medina, Mauricio Che, Yona F.A. Gaus, Samet Akcay, Toby P. Brekcon
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-01596

Browse the full library · About the Tower of Babel