FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability–Plasticity Tradeoff
Isaac Han Sangyeon Park Seungwon Oh Donghu Kim, Hojoon Lee* Kyung-Joong Kim* · 2026
Read FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability–Plasticity Tradeoff by Isaac Han Sangyeon Park Seungwon Oh Donghu Kim, Hojoon Lee* Kyung-Joong Kim* (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-2602-08040.txt.
Description
Neural networks learning from shifting data must stay stable enough to remember and plastic enough to adapt, and reinitialization methods struggle to balance the two. Isaac Han and five co-authors propose FIRE, which makes the trade-off plain and solvable.
Stability is measured by squared Frobenius error from past weights, and plasticity by deviation from isometry. The reinitialization point minimizes the first subject to the second being zero, approximated efficiently by Newton-Schulz iteration. Tests on continual image learning, language modeling, and reinforcement learning support the method. Continual-learning users gain a sound dial for the stability-plasticity balance instead of a hard-to-tune heuristic.
Catalog record
| CREATOR | Isaac Han Sangyeon Park Seungwon Oh Donghu Kim, Hojoon Lee* Kyung-Joong Kim* |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 71 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-08040