Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics

摘要: Numerous deep learning applications benefit from multi-task with multiple regression and classification objectives. In this paper we make the observation that performance of such systems is strongly dependent on relative weighting between each task's loss. Tuning these weights by hand a difficult expensive process, making prohibitive in practice. We propose principled approach to which weighs loss functions considering homoscedastic uncertainty task. This allows us simultaneously learn various quantities different units or scales both settings. demonstrate our model per-pixel depth regression, semantic instance segmentation monocular input image. Perhaps surprisingly, show can weightings outperform separate models trained individually