Direct-Sum Approach to Integrate Losses Via Classifier Subspace


Takumi Kobayashi (National Institute of Advanced Industrial Science and Technology (AIST))
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Deep models are successfully applied to various visual recognition tasks through end-to-end learning. A loss function is fundamental for the learning and various losses could be combined via arithmetic summation to improve performance. The simple summation, however, can bring about interference between the losses in back-propagation, deteriorating their synergy in the learning. In this paper, we propose a new approach to effectively integrate losses by mitigating the interference; we focus on classification and metric-based losses which are widely employed in discriminative supervised learning. The method leverages a classifier subspace to separate whole feature space into disjoint subspaces to which the two types of losses are respectively applied. Thereby, the losses are integrated in a direct-sum manner beyond a simple arithmetic summation to collaboratively work on learning feature representation without interference. In the experiments on few-shot image classification tasks which demand generalizable feature representation to unseen-class samples, the proposed method favorably improves performance by effectively combining the two types of losses.

Citation

@inproceedings{Kobayashi_2024_BMVC,
author    = {Takumi Kobayashi},
title     = {Direct-Sum Approach to Integrate Losses Via Classifier Subspace},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0900.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection