Generalized Greedy Gradient-Based Hyperparameter Optimization
Keywords: bilevel optimization, meta-learning, hyperparameter optimization
Abstract: Bilevel Optimization (BLO) is a widely-used approach that has numerous applications, including hyperparameter optimization, meta-learning. However, existing gradient-based method suffer from the following issues. Reverse-mode differentiation suffers from high memory requirements, while the methods based on the implicit function theorem require the convergence of the inner optimization. Approximations that consider a truncated inner optimization trajectory suffer from a short horizon bias. In this paper, we propose a novel approximation for hypergradient computation that sidesteps these difficulties. Specifically, we accumulate the short-horizon approximations from each step of the inner optimization trajectory. Additionally, we demonstrate that under certain conditions, the proposed hypergradient is a sufficient descent direction. Experimental results on a few-shot meta-learning and data hyper-cleaning tasks support our findings.
Primary Area: optimization
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.
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: 6297
Loading