StEVE: Adaptive Optimization in a Kronecker-Factored Eigenbasis

Jose Nicolas Marin Gamboa

Keywords: KFAC, EKFAC, Natural Gradient Descent, Adam, Optimization, Stochastic Optimization

TL;DR: Adam in the Kronecker-Factored Eigenbasis of the Empirical Fisher is faster than Adam, KFAC, and EKFAC

Abstract: Adaptive optimization algorithms such as Adam see widespread use in Deep Learning. However, these methods rely on diagonal approximations of the preconditioner, losing much information about the curvature of the loss surface and potentially leading to prolonged training times. We introduce StEVE (Stochastic Eigenbasis-adaptive Variance Estimation), a novel optimization algorithm that estimates lower order moments in the Kronecker-Factored Eigenbasis (KFE). By combining the advantages of Adam over other adaptive methods with the curvature-aware transformations of methods like KFAC and EKFAC, StEVE leverages second-order information while remaining computationally efficient. Our experiments demonstrate that EVE achieves faster convergence both in step-count and in wall-clock time compared to Adam, EKFAC, and KFAC for a variety of deep neural network architectures.

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.

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: 9031