Reproduce “Towards Few-Shot Adaptation of Foundation Models via Multitask Finetuning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhuoyan Xu, Zhenmei Shi, Junyi Wei, Fangzhou Mu, Yin Li, Yingyu Liang
A theoretical framework provides justification for multitask finetuning's efficacy in adapting pretrained foundation models to few-shot tasks, formally defining and quantifying how auxiliary task diversity and consistency reduce target task error. A practical greedy algorithm is introduced to select optimal subsets of auxiliary tasks based on these principles.
Key insights
- Multitask finetuning with a diverse and consistent set of auxiliary tasks provably reduces the generalization error on the target few-shot task compared to direct adaptation.
- The efficacy of multitask finetuning is quantitatively linked to auxiliary task ",diversity," (how broadly tasks cover target task characteristics) and ",consistency," (similarity between auxiliary and target task optimal models).
- The overall target task performance during finetuning is primarily determined by the total sample complexity (product of the number of auxiliary tasks and samples per task), rather than their individual values.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.