Optimizing ML Training with Metagradient Descent

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Engstrom, Logan, Ilyas, Andrew, Chen, Benjamin, Feldmann, Axel, Moses, William, Madry, Aleksander
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909540904075264
author Engstrom, Logan
Ilyas, Andrew
Chen, Benjamin
Feldmann, Axel
Moses, William
Madry, Aleksander
author_facet Engstrom, Logan
Ilyas, Andrew
Chen, Benjamin
Feldmann, Axel
Moses, William
Madry, Aleksander
contents A major challenge in training large-scale machine learning models is configuring the training process to maximize model performance, i.e., finding the best training setup from a vast design space. In this work, we unlock a gradient-based approach to this problem. We first introduce an algorithm for efficiently calculating metagradients -- gradients through model training -- at scale. We then introduce a "smooth model training" framework that enables effective optimization using metagradients. With metagradient descent (MGD), we greatly improve on existing dataset selection methods, outperform accuracy-degrading data poisoning attacks by an order of magnitude, and automatically find competitive learning rate schedules.
format Preprint
id arxiv_https___arxiv_org_abs_2503_13751
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Optimizing ML Training with Metagradient Descent
Engstrom, Logan
Ilyas, Andrew
Chen, Benjamin
Feldmann, Axel
Moses, William
Madry, Aleksander
Machine Learning
Artificial Intelligence
A major challenge in training large-scale machine learning models is configuring the training process to maximize model performance, i.e., finding the best training setup from a vast design space. In this work, we unlock a gradient-based approach to this problem. We first introduce an algorithm for efficiently calculating metagradients -- gradients through model training -- at scale. We then introduce a "smooth model training" framework that enables effective optimization using metagradients. With metagradient descent (MGD), we greatly improve on existing dataset selection methods, outperform accuracy-degrading data poisoning attacks by an order of magnitude, and automatically find competitive learning rate schedules.
title Optimizing ML Training with Metagradient Descent
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2503.13751