Gentle Local Robustness implies Generalization

Khoat Than, Dat Phan, Giang Vu · 2024

Read Gentle Local Robustness implies Generalization by Khoat Than, Dat Phan, Giang Vu (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-2412-06381.txt.

Description

Khoat Than, Dat Phan, and Giang Vu tighten the theory linking robustness to generalization in a paper from 2024. A model that shrugs off small input changes is expected to generalize. The authors show existing bounds fail a basic sanity test: they cannot approach the true error of the best possible classifier. That defect was previously unknown. Their new bounds depend on the model itself and provably converge to the ideal error as samples grow.

Experiments found two of the bounds informative for many deep networks pretrained on a large image set. Non-vacuous bounds are rare enough to matter. The work reframes how robustness certificates should be judged.

Catalog record

CREATORKhoat Than, Dat Phan, Giang Vu
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-06381

Browse the full library · About the Tower of Babel