Multi-task Curriculum Learning based on Gradient Similarity
Hiroaki Igarashi (DENSO Corporation),* Kenichi Yoneji (DENSO), Kohta Ishikawa (Denso IT Laboratory, Inc.), Rei Kawakami (Tokyo Institute of Technology), Teppei Suzuki (Denso IT Laboratory), Shingo Yashima (Denso IT Laboratory), Ikuro Sato (Tokyo Institute of Technology / Denso IT Laboratory)
The 33rd British Machine Vision Conference
Abstract
Intensive studies on multi-task learning (MTL) with deep neural networks have shown cases where both test error and computational cost can be reduced compared to single-task learning. However, several studies have argued that a naive implementation of MTL often degrades test performance due to gradient conflict, in which task-wise gradients have a negative inner product. These studies also invented ways to modify the gradients and eliminate the conflict. One concern about these methods is that the obtained solution is no longer optimal for the original objective due to the modification. In this paper, we propose a multi-task curriculum learning based on gradient similarity (MCLGS) to mitigate the negative impact of gradient conflicts while retaining the original objective toward the end of the training. We adopt a simple curriculum strategy that gives more weights to minibatches exhibiting fewer gradient conflicts in the early stage of training. We experimentally confirmed that MCLGS outperforms existing MTL methods, such as MGDA, PCGrad, GradDrop, and CAGrad, on BDD100K and NYUv2 datasets.Video
Citation
@inproceedings{Igarashi_2022_BMVC,
author = {Hiroaki Igarashi and Kenichi Yoneji and Kohta Ishikawa and Rei Kawakami and Teppei Suzuki and Shingo Yashima and Ikuro Sato},
title = {Multi-task Curriculum Learning based on Gradient Similarity},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0705.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by