arXiv 1705.07115
Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics
By Alex Kendall, Yarin Gal, et al.
Published 2017-05-19
Citation lineage
Review the prior work and downstream research connected to this paper.
Numerous deep learning applications benefit from multi-task learning with multiple regression and classification objectives. In this paper we make the observation that the performance of such systems is strongly dependent on the relative weighting between each task's loss. Tuning these weights by hand is a difficult and expensive process, making multi-task learning prohibitive in practice. We propose a principled ap…