MaxGNR: A Dynamic Weight Strategy via Maximizing Gradient-to-Noise Ratio for Multi-task Learning

When modeling related tasks in computer vision, Multi-Task Learning (MTL) can outperform Single-Task (STL) due to its ability capture intrinsic relatedness among tasks. However, MTL may encounter the insufficient training problem, i.e., some non-optimal situation compared with STL. A series of studies point out that too much gradient noise would lead performance degradation STL, however, scenario, Inter-Task Gradient Noise (ITGN) is an additional source for each task, which also affect optimization process. In this paper, we ITGN as a key factor leading problem. We define Gradient-to-Noise Ratio (GNR) measure relative magnitude and design MaxGNR algorithm alleviate interference task by maximizing GNR task. carefully evaluate our on two standard image datasets: NYUv2 Cityscapes. The results show outperforms baselines under identical experimental conditions.