HUB: Enhancing Learned Optimizers via Hybrid Update-based Strategy

Gaole Dai, Wei Wu, Ziyu Wang, Jie Fu, Shanghang Zhang, Tiejun Huang

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.

Keywords: learn-to-learn, leanred optimizer, hand-designed optimizer, optimization, meta learning, prompt tuning, large model

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Learned optimizers are pivotal in meta-learning and recent advancements in scalable learned optimizers have showcased superior performance over traditional, hand-designed counterparts in diverse tasks. However, their adoption is impeded by certain limitations, such as difficulties in handling out-of-distribution tasks, uncontrollable behaviors, and inferior performance in fine-tuning tasks. To address the issue of generalization in these optimizers, we propose a Hybrid-Update-Based (HUB) optimization strategy, inspired by the latest advancements in prompt tuning and result selection techniques in large language and vision models. Compared to previous methodologies (Pr'emont-Schwarz et al., 2022; Heaton et al., 2020), our approach enables a more sophisticated integration between hand-designed and learned optimizers and significantly reduces the computational overhead of hybridization. Our approach broadens the applicability of learned optimizers to tasks beyond their initial training distribution, and it has been validated through a series of diverse tasks, demonstrating significant advantages and unique robustness against out-of-distribution tasks compared to meticulously hyperparameter-tuned competitors. In this paper we also delve into a theoretical analysis of the hybrid strategy's impact on the behaviors and inherent traits of learned optimizers, offering deeper insights into their functionalities and interactions.

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: 652