Nash-GBML: Nash Gradient-Based Meta-Learning
Keywords: model-agnostic meta-learning, gradient-based meta-learning, game theory, nash game, single-leader multi-follower game
TL;DR: We propose the Nash gradient-based meta-learning which considers the interaction among tasks and model it as the single-leader multi-follower game.
Abstract: Meta-learning has been proposed to address fast adaptation to unseen tasks with little data. Traditional meta-learning is modeled as the Single-Leader Multi-Follower game consisting of inner and outer-level problems to minimize average or worst-case task loss. Because they assume all sampled tasks are independent, it reduces the flexibility of modeling complex interaction among tasks. Thus, we formulate meta-learning as a Single-Leader Multi-Follower game by considering the interaction among tasks at the inner level. We propose the Nash-GBML incorporating a penalty term into the task loss function to model the interaction among task-specific parameters. We discuss the iteration complexity and convergence of the Nash-GBML algorithm. To validate our Nash-GBML algorithm, we introduce two penalty terms, which are designed to reduce the average and worst-case task loss. We empirically show that the Nash-GBML with the proposed penalty terms outperforms traditional GBML for supervised learning experiments.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2844
Loading