Chentong Huang

Regularity-Aware Stochastic MGDA with Adaptive Conflict-Avoidant Update Direction Control

Chentong Huang, Lisha Chen

Multi-objective learning (MOL) aims to optimize multiple objectives simultaneously. The multi-gradient descent algorithm (MGDA) is a workhorse that iteratively updates along a common descent or conflict-avoidant (CA) direction across objectives. In stochastic settings, however, t…Read more