Towards Robust Influence Functions with Flat Validation Minima
Xichen Ye, Yifan Wu, Weizhong Zhang, Cheng Jin, Yifan Chen · 2025
Read Towards Robust Influence Functions with Flat Validation Minima by Xichen Ye, Yifan Wu, Weizhong Zhang, Cheng Jin, Yifan Chen (2025) 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-2505-19097.txt.
Description
Xichen Ye and four co-authors work on making influence estimates more trustworthy in deep learning in this 2025 paper. Influence functions gauge how much each training example sways a model's answers. In large networks, those estimates often fail, above all with noisy training data.
The authors trace the fault not to how parameter change is estimated, but to how loss change is estimated. Sharp dips in the check-set loss make those estimates unreliable. They prove a link between estimation error, check-set risk, and the sharpness of that risk. They then design a new influence formula built for flat, broad low spots in the check loss. Tests across several tasks showed their method beat the standard approach.
Catalog record
| CREATOR | Xichen Ye, Yifan Wu, Weizhong Zhang, Cheng Jin, Yifan Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-19097