Degradation of Feature Space in Continual Learning

Chiara Lanza 1 Roberto Pereira 1 Marco Miozzo 1, Eduard Angelats 2 Paolo Dini 1 · 2026

Read Degradation of Feature Space in Continual Learning by Chiara Lanza 1 Roberto Pereira 1 Marco Miozzo 1, Eduard Angelats 2 Paolo Dini 1 (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-2602-06586.txt.

Description

Continual learning trains models on streaming, shifting data, which forces a trade-off between learning new tasks and keeping old ones. Chiara Lanza and 4 co-authors ask whether enforcing feature-space evenness, the even spread of features that helps one-shot training, could ease that trade-off. In one-shot training, even features support structured, usable features almost by default.

Their tests used contrast-based continual learning on two standard image tests. Even-feature penalty did not help; it could actually degrade scores in continual settings. The finding highlights a real shape gap between one-shot and continual learning. Evenness, beneficial when data arrive all at once, may be the wrong built-in bias when data arrive as a shifting stream, and the authors suggest the field should look elsewhere for balance.

Catalog record

CREATORChiara Lanza 1 Roberto Pereira 1 Marco Miozzo 1, Eduard Angelats 2 Paolo Dini 1
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-06586

Browse the full library · About the Tower of Babel