Mirror, Mirror of the Flow: How Does Regularization Shape Implicit Bias?

Tom Jacobs, Chao Zhou, Rebekka Burkholz

Keywords: Implicit bias, explicit regularization, weight decay, matrix sensing, LoRA, attention, mirror flow, time-dependent Legendre function

TL;DR: This work studies how explicit regularization influences implicit bias in overparameterized models, with applications spanning various problem domains, including sparse coding, matrix sensing, single-layer attention and LoRA.

Abstract: Implicit bias plays an important role in explaining how overparameterized models generalize well. Explicit regularization like weight decay is often employed in addition to prevent overfitting. While both concepts have been studied separately, in practice, they often act in tandem. Understanding their interplay is key to controlling the shape and strength of implicit bias, as it can be modified by explicit regularization. To this end, we incorporate explicit regularization into the mirror flow framework and analyze its lasting effects on the geometry of the training dynamics, covering three distinct effects: positional bias, type of bias, and range shrinking. Our analytical approach encompasses a broad class of problems, including sparse coding, matrix sensing, single-layer attention, and LoRA, for which we demonstrate the utility of our insights. To exploit the lasting effect of regularization and highlight the potential benefit of dynamic weight decay schedules, we propose to switch off weight decay during training, which can improve generalization, as we demonstrate in experiments.

Submission Number: 20