Model Merging by Gradient Matching

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.

Cite

Text

Daheim et al. "Model Merging by Gradient Matching." NeurIPS 2023 Workshops: UniReps, 2023.

Markdown

[Daheim et al. "Model Merging by Gradient Matching." NeurIPS 2023 Workshops: UniReps, 2023.](https://mlanthology.org/neuripsw/2023/daheim2023neuripsw-model/)

BibTeX

@inproceedings{daheim2023neuripsw-model,
  title     = {{Model Merging by Gradient Matching}},
  author    = {Daheim, Nico and Möllenhoff, Thomas and Ponti, Edoardo and Gurevych, Iryna and Khan, Mohammad Emtiyaz},
  booktitle = {NeurIPS 2023 Workshops: UniReps},
  year      = {2023},
  url       = {https://mlanthology.org/neuripsw/2023/daheim2023neuripsw-model/}
}