Injecting Imbalance Sensitivity for Multi-Task Learning
This addresses imbalance issues in multi-task learning for AI applications, though it appears incremental as it enhances an existing baseline method.
The paper tackles the problem of imbalance/dominance in multi-task learning, which previous gradient-based methods had neglected, by proposing an IMbalance-sensitive Gradient (IMGrad) descent method that imposes constraints on projected norms. The method demonstrates competitive performance on multiple mainstream MTL benchmarks in supervised and reinforcement learning tasks.
Multi-task learning (MTL) has emerged as a promising approach for deploying deep learning models in real-life applications. Recent studies have proposed optimization-based learning paradigms to establish task-shared representations in MTL. However, our paper empirically argues that these studies, specifically gradient-based ones, primarily emphasize the conflict issue while neglecting the potentially more significant impact of imbalance/dominance in MTL. In line with this perspective, we enhance the existing baseline method by injecting imbalance-sensitivity through the imposition of constraints on the projected norms. To demonstrate the effectiveness of our proposed IMbalance-sensitive Gradient (IMGrad) descent method, we evaluate it on multiple mainstream MTL benchmarks, encompassing supervised learning tasks as well as reinforcement learning. The experimental results consistently demonstrate competitive performance.