Work

Training Deep Learning Models with Norm-Constrained LMOs

Year: 2025

Type: preprint

Abstract: In this work, we study optimization methods that leverage the linear minimization oracle (LMO) over a norm-ball. We propose a new stochastic family of algorithms that uses the LMO to adapt to the geometry of the problem and, perhaps surprisingly, show that they can be applied to unconstrained problems. The resulting update rule unifies several existing optimization methods under a single framework. Furthermore, we propose an explicit choice of norm for deep architectures, which, as a side benefit, leads to the transferability of hyperparameters across model sizes. Experimentally, we demonstrate significant speedups on nanoGPT training using our algorithm, Scion, without any reliance on Adam. The proposed method is memory-efficient, requiring only one set of model weights and one set of gradients, which can be stored in half-precision. The code is available at https://github.com/LIONS-EPFL/scion .

Source: arXiv (Cornell University)

DOI: 10.48550/arxiv.2502.07529

Authors: Pethick, Thomas, Xie, Wanyun, Antonakopoulos, Kimon, Zhu, Zhenyu, Silveti-Falls, Antonio, Cevher, Volkan

Institutions: École Polytechnique Fédérale de Lausanne, Université Paris-Saclay, CentraleSupélec, Centre Inria de Saclay, Centre de Vision Numérique

Language: English

Topic: Stochastic Gradient Optimization Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

Funders: National Science Foundation, Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung, Hasler Stiftung, Fondation Mathématique Jacques Hadamard, Centro Svizzero di Calcolo Scientifico, National Supercomputing Centre Singapore, Army Research Office

Awards: G7452299184, Deep Optimization, G993971353

OpenAlex ID: W4407424967

Cites:

Cited by:

Related to: 10