Hybrid Least Squares/Gradient Descent Methods for DeepONets

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Choi, Jun, Lee, Chang-Ock, Moon, Minam
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916911171764224
author Choi, Jun
Lee, Chang-Ock
Moon, Minam
author_facet Choi, Jun
Lee, Chang-Ock
Moon, Minam
contents We propose an efficient hybrid least squares/gradient descent method to accelerate DeepONet training. Since the output of DeepONet can be viewed as linear with respect to the last layer parameters of the branch network, these parameters can be optimized using a least squares (LS) solve, and the remaining hidden layer parameters are updated by means of gradient descent form. However, building the LS system for all possible combinations of branch and trunk inputs yields a prohibitively large linear problem that is infeasible to solve directly. To address this issue, our method decomposes the large LS system into two smaller, more manageable subproblems $\unicode{x2014}$ one for the branch network and one for the trunk network $\unicode{x2014}$ and solves them separately. This method is generalized to a broader type of $L^2$ loss with a regularization term for the last layer parameters, including the case of unsupervised learning with physics-informed loss.
format Preprint
id arxiv_https___arxiv_org_abs_2508_15394
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Hybrid Least Squares/Gradient Descent Methods for DeepONets
Choi, Jun
Lee, Chang-Ock
Moon, Minam
Machine Learning
Artificial Intelligence
Numerical Analysis
47-08, 65F45, 65Y10, 68T07, 68T20
We propose an efficient hybrid least squares/gradient descent method to accelerate DeepONet training. Since the output of DeepONet can be viewed as linear with respect to the last layer parameters of the branch network, these parameters can be optimized using a least squares (LS) solve, and the remaining hidden layer parameters are updated by means of gradient descent form. However, building the LS system for all possible combinations of branch and trunk inputs yields a prohibitively large linear problem that is infeasible to solve directly. To address this issue, our method decomposes the large LS system into two smaller, more manageable subproblems $\unicode{x2014}$ one for the branch network and one for the trunk network $\unicode{x2014}$ and solves them separately. This method is generalized to a broader type of $L^2$ loss with a regularization term for the last layer parameters, including the case of unsupervised learning with physics-informed loss.
title Hybrid Least Squares/Gradient Descent Methods for DeepONets
topic Machine Learning
Artificial Intelligence
Numerical Analysis
47-08, 65F45, 65Y10, 68T07, 68T20
url https://arxiv.org/abs/2508.15394