Multi-task gradient descent for multi-task learning

• Main Authors: Bai, Lu, Ong, Yew-Soon, He, Tiantian, Gupta, Abhishek
• Other Authors: School of Computer Science and Engineering
• Format: Article
• Language: English
• Published: 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Knowledge Transfer; Cost Functions
• Online Access: https://hdl.handle.net/10356/147806
• Institution: Nanyang Technological University

Multi-Task Learning (MTL) aims to simultaneously solve a group of related learning tasks by leveraging the salutary knowledge memes contained in the multiple tasks to improve the generalization performance. Many prevalent approaches focus on designing a sophisticated cost function, which integrates all the learning tasks and explores the task-task relationship in a predefined manner. Different from previous approaches, in this paper, we propose a novel Multi-task Gradient Descent (MGD) framework, which improves the generalization performance of multiple tasks through knowledge transfer. The uniqueness of MGD lies in assuming individual task-specific learning objectives at the start, but with the cost functions implicitly changing during the course of parameter optimization based on task-task relationships. Specifically, MGD optimizes the individual cost function of each task using a reformative gradient descent iteration, where relations to other tasks are facilitated through effectively transferring parameter values (serving as the computational representations of memes) from other tasks. Theoretical analysis shows that the proposed framework is convergent under any appropriate transfer mechanism. Compared with existing MTL approaches, MGD provides a novel easy-to-implement framework for MTL, which can mitigate negative transfer in the learning procedure by asymmetric transfer. The proposed MGD has been compared with both classical and state-of-the-art approaches on multiple MTL datasets. The competitive experimental results validate the effectiveness of the proposed algorithm.