A Model-Agnostic Approach to Mitigate Gradient Interference for Multi-Task Learning
提出一种模型无关的梯度裁剪方法MAMG，通过直接修改梯度干扰方向上的干扰分量来缓解多任务学习中的任务冲突，理论证明收敛性，实验表明优于现有方法。
Multitask learning (MTL) is a powerful technique for jointly learning multiple tasks. However, it is difficult to achieve a tradeoff between tasks during iterative training, as some tasks may compete with each other. Existing methods manually design specific network models to mitigate task conflicts, but they require considerable manual effort and prior knowledge about task relationships to tune the model so as to obtain the best performance for each task. Moreover, few works have offered formal descriptions of task conflicts and theoretical explanations for the cause of task conflict problems. In this article, we provide a formal description of task conflicts that are caused by the gradient interference problem of tasks. To alleviate this issue, we propose a novel model-agnostic approach to mitigate gradient interference (MAMG) by designing a gradient clipping rule that directly modifies the interfering components on the gradient interfering direction. Specifically, MAMG is model-agnostic and thus it can be applied to a large number of multitask models. We also theoretically prove the convergence of MAMG and its superiority to existing MTL methods. We evaluate our method on a variety of real-world large datasets, and extensive experimental results confirm that MAMG can outperform some state-of-the-art algorithms on different types of tasks and can be easily applied to various methods.