How to Weight Multitask Finetuning? Fast Previews via Bayesian Model-Merging
This work addresses the challenge of costly weight search in multitask finetuning for machine learning practitioners, offering an incremental improvement through Bayesian enhancements to model merging.
The paper tackles the problem of efficiently weighting multiple tasks during finetuning by proposing fast previews using Bayesian model-merging, which reuses and averages parameters from separately trained models without retraining. The approach is validated on vision and natural-language transformers, showing improved multitask finetuning performance.
When finetuning multiple tasks altogether, it is important to carefully weigh them to get a good performance, but searching for good weights can be difficult and costly. Here, we propose to aid the search with fast previews to quickly get a rough idea of different reweighting options. We use model merging to create previews by simply reusing and averaging parameters of models trained on each task separately (no retraining required). To improve the quality of previews, we propose a Bayesian approach to design new merging strategies by using more flexible posteriors. We validate our findings on vision and natural-language transformers. Our work shows the benefits of model merging via Bayes to improve multitask finetuning.