MetaWeighting: Learning to Weight Tasks in Multi-Task Learning

Findings of the Association for Computational Linguistics: ACL 2022, pages 3436 - 3448 May 22-27, 2022 c⃝2022 Association for Computational Linguistics MetaWeighting: Learning to Weight Tasks in Multi-Task Learning Yuren Mao1, Zekai Wang2, Weiwei Liu2∗, Xuemin Lin1, Pengtao Xie3 1School of Computer Science and Engineering, University of New South Wales 2School of Computer Science, Wuhan University 3Department of Electrical and Computer Engineering, University of California San Diego yuren.mao@unsw.edu.au, {wzekai99,liuweiwei863}@gmail.com lxue@cse.unsw.edu.au, pengtaoxie2008@gmail.com Abstract Task weighting, which assigns weights on the including tasks during training, significantly matters the performance of Multi-task Learn- ing (MTL); thus, recently, there has been an explosive interest in it. However, existing task weighting methods assign weights only based on the training loss, while ignoring the gap between the training loss and generalization loss. It degenerates MTL’s performance. To address this issue, the present paper proposes a novel task weighting algorithm, which automat- ically weights the tasks via a learning-to-learn paradigm, referred to as MetaWeighting. Ex- tensive experiments are conducted to validate the superiority of our proposed method in multi- task text classification. 1 Introduction Multi-task Learning (MTL) simultaneously learns multiple related tasks and aims to achieve better performance than learning each task independently (Caruana, 1993; Baxter, 2000). It has achieved great success in various applications; especially, in the text classification context, MTL can signifi- cantly outperform single task learning (Liu et al., 2017; Mao et al., 2021). In MTL, it is common for the including tasks to be competing. If we cannot properly balance these tasks, some tasks might dominate the training process and hurt the performance of other tasks, a phenomenon known as task imbalance. To address the task imbalance, the most widely used method is task weighting, which adaptively assigns weights on the tasks during training to balance their im- pacts. Various task weighting methods have been proposed and can be used in multi-task text classi- fication, such as (Kendall et al., 2018; Sener and Koltun, 2018; Chen et al., 2018). However, existing task weighting methods com- pute the task weights only based on training losses or corresponding gradients. They ignore the gap Figure 1: Illustration of the gap between training loss and generalization loss in the training process of a four- task topic classification experiment (500th , 1000th, 1500th epochs respectively). between the training loss and generalization loss. To illustrate this gap, we report observations of our four-task topic classification experiment in Figure 1. The detailed experimental settings are intro- duced in the experiment section. Figure 1 demon- strates that the training losses and generalization losses (estimated by the test losses) have different magnitudes; moreover, they have different patterns, such as a task might have the largest training loss but the lowest generalization loss among the tasks. This gap causes a mismatch between the task weights and tasks’ generalization performance, which reduces effectiveness of the task weighting. To tackle this issue, this paper proposes a novel task weighting method based on a bi-level optimization problem, which aims to find task weights that ex- plicitly …

Content truncated to 500 words