Salvato in:
Dettagli Bibliografici
Autori principali: Fernando, Heshan, Shen, Han, Liu, Miao, Chaudhury, Subhajit, Murugesan, Keerthiram, Chen, Tianyi
Natura: Preprint
Pubblicazione: 2022
Soggetti:
Accesso online:https://arxiv.org/abs/2210.12624
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866916164645421056
author Fernando, Heshan
Shen, Han
Liu, Miao
Chaudhury, Subhajit
Murugesan, Keerthiram
Chen, Tianyi
author_facet Fernando, Heshan
Shen, Han
Liu, Miao
Chaudhury, Subhajit
Murugesan, Keerthiram
Chen, Tianyi
contents Machine learning problems with multiple objective functions appear either in learning with multiple criteria where learning has to make a trade-off between multiple performance metrics such as fairness, safety and accuracy; or, in multi-task learning where multiple tasks are optimized jointly, sharing inductive bias between them. This problems are often tackled by the multi-objective optimization framework. However, existing stochastic multi-objective gradient methods and its variants (e.g., MGDA, PCGrad, CAGrad, etc.) all adopt a biased noisy gradient direction, which leads to degraded empirical performance. To this end, we develop a stochastic Multi-objective gradient Correction (MoCo) method for multi-objective optimization. The unique feature of our method is that it can guarantee convergence without increasing the batch size even in the non-convex setting. Simulations on multi-task supervised and reinforcement learning demonstrate the effectiveness of our method relative to state-of-the-art methods.
format Preprint
id arxiv_https___arxiv_org_abs_2210_12624
institution arXiv
publishDate 2022
record_format arxiv
spellingShingle Mitigating Gradient Bias in Multi-objective Learning: A Provably Convergent Stochastic Approach
Fernando, Heshan
Shen, Han
Liu, Miao
Chaudhury, Subhajit
Murugesan, Keerthiram
Chen, Tianyi
Machine Learning
Optimization and Control
Machine learning problems with multiple objective functions appear either in learning with multiple criteria where learning has to make a trade-off between multiple performance metrics such as fairness, safety and accuracy; or, in multi-task learning where multiple tasks are optimized jointly, sharing inductive bias between them. This problems are often tackled by the multi-objective optimization framework. However, existing stochastic multi-objective gradient methods and its variants (e.g., MGDA, PCGrad, CAGrad, etc.) all adopt a biased noisy gradient direction, which leads to degraded empirical performance. To this end, we develop a stochastic Multi-objective gradient Correction (MoCo) method for multi-objective optimization. The unique feature of our method is that it can guarantee convergence without increasing the batch size even in the non-convex setting. Simulations on multi-task supervised and reinforcement learning demonstrate the effectiveness of our method relative to state-of-the-art methods.
title Mitigating Gradient Bias in Multi-objective Learning: A Provably Convergent Stochastic Approach
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2210.12624