Separated and Independent Contrastive Learning on Labeled and Unlabeled Samples: Boosting Performance on Long-tail Semi-supervised Learning


Dongyoung Kim (Hallym University), Jeong-Gun Lee (Hallym University), WonSook Lee (University of Ottawa)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

Conventional semi-supervised learning (SSL) encounters challenges in effectively addressing issues associated with long-tail problems, primarily stemming from imbalances within a dataset. Previous semi-supervised approaches incorporating contrastive learning relied on unlabeled samples to apply the contrastive learning method. Consequently, to identify positive samples from unlabeled ones, they needed to make pseudo-labels, but inaccurate pseudo-labels lead to confirmation bias toward majority classes in long-tail datasets. Therefore, we want to obtain meaningful information from labeled samples which include accurate labels. In this paper, we introduce Seperated Independent Contrastive Semi-Supervised Learning (SICSSL) for long-tail, which leverages a supervised contrastive learning approach for labeled samples and unlabeled samples separately and independently to enhance performance. In our experiments, employing labeled samples for contrastive learning yields superior performance compared to the contrastive learning using only unlabeled samples.

Citation

@inproceedings{Kim_2024_BMVC,
author    = {Dongyoung Kim and Jeong-Gun Lee and WonSook Lee},
title     = {Separated and Independent Contrastive Learning on Labeled and Unlabeled Samples: Boosting Performance on Long-tail Semi-supervised Learning},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0433.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