Conflict-Minimizing Gradient Selection for Multi-Task Learning

Machine Learning (ML) models are typically trained on data for a single target task, with the goal of optimizing performance for that specific task. In contrast, multi-task learning (MTL) aims to train a single model on data from multiple tasks simultaneously, where a shared model representation is learned across all tasks. MTL is widely applied across various domains, including natural language processing and computer vision, but training multiple tasks together introduces significant challenges.
MTL is a multi-objective optimization problem, different tasks often have competing objectives, leading to conflicts in gradient updates. Conflict management is critical in this case to facilitate effective learning. A key challenge lies in selecting an optimal update direction that balances the trade-offs between competing tasks without compromising overall model performance.
This study presents a comparative analysis of methods for handling gradient interference in MTL. It examines different strategies for selecting a single optimal direction in multi-objective optimization. The algorithms are tested on computer vision tasks using the CityScapes and NYU-v2 datasets and the GLUE benchmark for natural language processing, providing a comprehensive evaluation of the effectiveness of these approaches.