Method Details

Details for method 'Multitask Learning'

name	Multitask Learning
challenge	instance-level semantic labeling
details	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 approach to multi-task deep learning which weighs multiple loss functions by considering the homoscedastic uncertainty of each task. This allows us to simultaneously learn various quantities with different units or scales in both classification and regression settings. We demonstrate our model learning per-pixel depth regression, semantic and instance segmentation from a monocular input image. Perhaps surprisingly, we show our model can learn multi-task weightings and outperform separate models trained individually on each task.
publication	Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics Alex Kendall, Yarin Gal and Roberto Cipolla CVPR 2018 https://arxiv.org/abs/1705.07115
project page / code	https://github.com/alexgkendall/multitaskvision
used Cityscapes data	fine annotations
used external data	ImageNet
runtime	n/a
subsampling	no
submission date	April, 2018
previous submissions

Class	AP	AP50%	AP100m	AP50m
person	19.2287	38.1028	38.1629	38.8931
rider	21.3894	46.2562	35.5064	36.3393
car	36.5683	54.7505	60.4266	64.0127
truck	18.7961	28.4405	26.6905	29.0803
bus	26.8213	40.778	42.6905	46.6384
train	15.8773	25.0182	24.4086	28.188
motorcycle	19.3886	42.2056	27.4564	28.5165
bicycle	14.5134	36.5269	24.6618	24.6738