Time-Dependent Mirror Flows and Where to Find Them
Keywords: Mirror flow, Implicit Bias, Time-dependent Bregman potential, Explicit regularization, LoRA, Attention, Sparse coding
Abstract: Explicit regularization and implicit bias are often studied separately, though in practice, they act in tandem. However, their interplay remains poorly understood. In this work, we show that explicit regularization modifies the behavior of implicit bias and provides a mechanism to control its strength. By incorporating explicit regularization into the mirror flow framework, we present a general approach to better understand implicit biases and their potential in guiding the design of optimization problems. Our primary theoretical contribution is the characterization of regularizations and reparameterizations that induce a time-dependent Bregman function, with a discussion of the implications of its temporal variation. Importantly, our framework encompasses single-layer attention, and application to sparse coding. Extending beyond our core assumptions, we apply this framework to LoRA finetuning, revealing an implicit bias towards sparsity.
Supplementary Material: zip
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: 4915
Loading