Gated subspace alignment with drift compensation for parameter-efficient Class-Incremental Learning
Jianye Gu, Shucheng Huang, Tian Li, Senbao Zhang, Mingxing Li · 2026
Read Gated subspace alignment with drift compensation for parameter-efficient Class-Incremental Learning by Jianye Gu, Shucheng Huang, Tian Li, Senbao Zhang, Mingxing Li (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/plos-pone-0348270.txt.
Description
Jianye Gu and co-workers tackle a stubborn problem for learning machines: adding new categories over time without forgetting old ones. As models expand their inner feature spaces, memories of earlier classes drift and blur. Their 2026 method freezes the core network and adds tiny gated adapters for each new task, building separate subspaces that still share information through attention.
A learned mapping then lines old class memories up inside the newest space, all without revisiting old training data. Benchmark tests showed better accuracy and stronger resistance to forgetting than rival approaches. Keeping the core frozen also kept computing costs low. The design suits tools that must keep learning on the job, from photo sorting to medical screening.
Catalog record
| CREATOR | Jianye Gu, Shucheng Huang, Tian Li, Senbao Zhang, Mingxing Li |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65266 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |