AdaTask: A Task-aware Adaptive Learning Rate Approach to Multi-task Learning
Enneng Yang, Junwei Pan, Ximei Wang, Haibin Yu, Li Shen, Xihua Chen, Lei Xiao, Jie Jiang, Guibing Guo
TL;DR
AdaTask tackles task dominance in multi-task learning by introducing per-task accumulative gradients and an AU/rAU-based metric to quantify parameter-wise task influence. By separating accumulative gradients for each task within adaptive optimizers (e.g., AdaGrad, RMSProp, Adam), AdaTask prevents any single task from dominating learning rates or updates, leading to improved performance on dominated tasks while preserving strong overall metrics. Extensive experiments across CityScapes, TikTok, and WeChat demonstrate substantial gains on dominated tasks and competitive or state-of-the-art average task performance, with additional gains achievable when combined with gradient-direction methods. The work also provides analysis showing AdaTask balances shared parameters across layers and reduces learning-rate dominance, offering a practical, architecture-agnostic approach to robust multi-task optimization.
Abstract
Multi-task learning (MTL) models have demonstrated impressive results in computer vision, natural language processing, and recommender systems. Even though many approaches have been proposed, how well these approaches balance different tasks on each parameter still remains unclear. In this paper, we propose to measure the task dominance degree of a parameter by the total updates of each task on this parameter. Specifically, we compute the total updates by the exponentially decaying Average of the squared Updates (AU) on a parameter from the corresponding task.Based on this novel metric, we observe that many parameters in existing MTL methods, especially those in the higher shared layers, are still dominated by one or several tasks. The dominance of AU is mainly due to the dominance of accumulative gradients from one or several tasks. Motivated by this, we propose a Task-wise Adaptive learning rate approach, AdaTask in short, to separate the \emph{accumulative gradients} and hence the learning rate of each task for each parameter in adaptive learning rate approaches (e.g., AdaGrad, RMSProp, and Adam). Comprehensive experiments on computer vision and recommender system MTL datasets demonstrate that AdaTask significantly improves the performance of dominated tasks, resulting SOTA average task-wise performance. Analysis on both synthetic and real-world datasets shows AdaTask balance parameters in every shared layer well.