Skip to yearly menu bar Skip to main content


Poster

Model Merging by Uncertainty-Based Gradient Matching

Nico Daheim · Thomas Möllenhoff · Edoardo M. Ponti · Iryna Gurevych · Mohammad Emtiyaz Khan

Halle B #221
[ ]
Tue 7 May 1:45 a.m. PDT — 3:45 a.m. PDT

Abstract:

Models trained on different datasets can be merged by a weighted-averaging of their parameters, but why does it work and when can it fail? Here, we connect the inaccuracy of weighted-averaging to mismatches in the gradients and propose a new uncertainty-based scheme to improve the performance by reducing the mismatch. The connection also reveals implicit assumptions in other schemes such as averaging, task arithmetic, and Fisher-weighted averaging. Our new method gives consistent improvements for large language models and vision transformers, both in terms of performance and robustness to hyperparameters.

Chat is not available.