vix.ing · top · new · best · stats · spec

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

2017/05/19 by Alex Kendall, Yarin Gal, Kendall, Alex +3 · 182 citations
Computer Science · #Domain Adaptation and Few-Shot Learning #Advanced Neural Network Applications #Human Pose and Action Recognition

paper · pdf · doi:10.48550/arxiv.1705.07115

Abstract

Numerous deep learning applications benefit from multi-task learning with\nmultiple regression and classification objectives. In this paper we make the\nobservation that the performance of such systems is strongly dependent on the\nrelative weighting between each task's loss. Tuning these weights by hand is a\ndifficult and expensive process, making multi-task learning prohibitive in\npractice. We propose a principled approach to multi-task deep learning which\nweighs multiple loss functions by considering the homoscedastic uncertainty of\neach task. This allows us to simultaneously learn various quantities with\ndifferent units or scales in both classification and regression settings. We\ndemonstrate our model learning per-pixel depth regression, semantic and\ninstance segmentation from a monocular input image. Perhaps surprisingly, we\nshow our model can learn multi-task weightings and outperform separate models\ntrained individually on each task.\n

Cited by

Related