Hessian Surgery: Class-Targeted Post-Hoc Rebalancing via Hessian Spike Perturbation
Hugo Vigna Samuel Bontemps CentraleSupélec – Université Paris-Saclay ESILV – Léonard de Vinci [email protected] [email protected] · 2026
Read Hessian Surgery: Class-Targeted Post-Hoc Rebalancing via Hessian Spike Perturbation by Hugo Vigna Samuel Bontemps CentraleSupélec – Université Paris-Saclay ESILV – Léonard de Vinci [email protected] [email protected] (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-2605-07790.txt.
Description
Hugo Vigna and Samuel Bontemps perform surgery on trained neural networks to fix their weakest classes, without retraining. The Hessian of a trained network has a telling spectrum: a bulk of near-zero eigenvalues plus a few large spikes, as many as the number of classes minus one. Earlier work described this structure but never used it operationally.
Their Hessian Surgery perturbs weights along the spike eigenvectors to rebalance per-class accuracy. A sensitivity matrix measures how each class's accuracy responds along each spike direction, a constrained optimization targets weak classes while protecting strong ones, and an adaptive budget grows or shrinks with progress. On CIFAR-10 and the ISIC-2019 skin lesion dataset, the method improved balanced accuracy and cut its spread across classes.
Catalog record
| CREATOR | Hugo Vigna Samuel Bontemps CentraleSupélec – Université Paris-Saclay ESILV – Léonard de Vinci [email protected] [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 86 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.07790