Enregistré dans:
Détails bibliographiques
Auteurs principaux: Jin, Lisa, Ma, Jianhao, Liu, Zechun, Gromov, Andrey, Defazio, Aaron, Xiao, Lin
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2503.15748
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917963182899200
author Jin, Lisa
Ma, Jianhao
Liu, Zechun
Gromov, Andrey
Defazio, Aaron
Xiao, Lin
author_facet Jin, Lisa
Ma, Jianhao
Liu, Zechun
Gromov, Andrey
Defazio, Aaron
Xiao, Lin
contents We develop a principled method for quantization-aware training (QAT) of large-scale machine learning models. Specifically, we show that convex, piecewise-affine regularization (PAR) can effectively induce the model parameters to cluster towards discrete values. We minimize PAR-regularized loss functions using an aggregate proximal stochastic gradient method (AProx) and prove that it has last-iterate convergence. Our approach provides an interpretation of the straight-through estimator (STE), a widely used heuristic for QAT, as the asymptotic form of PARQ. We conduct experiments to demonstrate that PARQ obtains competitive performance on convolution- and transformer-based vision tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2503_15748
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle PARQ: Piecewise-Affine Regularized Quantization
Jin, Lisa
Ma, Jianhao
Liu, Zechun
Gromov, Andrey
Defazio, Aaron
Xiao, Lin
Machine Learning
Optimization and Control
We develop a principled method for quantization-aware training (QAT) of large-scale machine learning models. Specifically, we show that convex, piecewise-affine regularization (PAR) can effectively induce the model parameters to cluster towards discrete values. We minimize PAR-regularized loss functions using an aggregate proximal stochastic gradient method (AProx) and prove that it has last-iterate convergence. Our approach provides an interpretation of the straight-through estimator (STE), a widely used heuristic for QAT, as the asymptotic form of PARQ. We conduct experiments to demonstrate that PARQ obtains competitive performance on convolution- and transformer-based vision tasks.
title PARQ: Piecewise-Affine Regularized Quantization
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2503.15748