Model Merging by Uncertainty-Based Gradient Matching
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...
arxiv.org