Mitigating Forgetting in Continual Learning with Selective Gradient Projection

Anika Singh, Aayush Dhaulakhandi, Varun Chopade, Likhith Malipati, Kevin Zhu · 2026

Read Mitigating Forgetting in Continual Learning with Selective Gradient Projection by Anika Singh, Aayush Dhaulakhandi, Varun Chopade, Likhith Malipati, Kevin Zhu (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-2603-26671.txt.

Description

Singh and co-authors present Selective Forgetting-Aware Optimization, a continual learning method that guards old skills while new ones form. Networks tend to overwrite past knowledge when they adapt, a failure called catastrophic forgetting. Their method steers gradient directions by cosine similarity and per-layer gates. Updates can be projected, accepted, or dropped under a tunable rule, with a Monte Carlo shortcut keeping it cheap. On standard benchmarks it matches rival accuracy while cutting memory cost by 90%, suiting tight devices.

Catalog record

CREATORAnika Singh, Aayush Dhaulakhandi, Varun Chopade, Likhith Malipati, Kevin Zhu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44846
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.26671

Browse the full library · About the Tower of Babel