A Balanced Multi-Task Learning Method for Efficient Task Optimization*
Xiaohan Zhong, Xinyan Lu, Jinxia Zhang, Shixiong Fang · 2025
Multi-task learning effectively addresses related tasks by leveraging shared knowledge while optimizing taskspecific objectives. However, Multi-task learning often restricts from task conflicts and the “seesaw” effect, where improving one task negatively impacts another. To solve these issues, we propose the Adaptive Progressive Learning Module, which dynamically balances shared and task-specific learning through Specific Experts, Shared Experts, and Gate Network. Our module enhances knowledge sharing while mitigating task interference and imbalances. Evaluated on the Cityscapes and NYU Depth v2 datasets, our methods demonstrate superior performance compared to state-of-the-art methods. It achieves balanced task optimization, significantly reducing the “seesaw” effect. The proposed framework provides a scalable and flexible solution for multi-task learning, offering robust performance across tasks with varying correlations.