Login (DCU Staff Only)
Login (DCU Staff Only)

DORAS | DCU Research Repository

Explore open access research and scholarly works from DCU

Advanced Search

Unsupervised label noise modeling and loss correction

Arazo Sánchez, Eric, Ortego, Diego orcid logoORCID: 0000-0002-1011-3610, Albert, Paul, O'Connor, Noel E. orcid logoORCID: 0000-0002-4033-9135 and McGuinness, Kevin orcid logoORCID: 0000-0003-1336-6477 (2019) Unsupervised label noise modeling and loss correction. In: International Conference on Machine Learning (ICML), 10-15 June 2019, Long Beach, CA, USA.

Abstract
Despite being robust to small amounts of label noise, convolutional neural networks trained with stochastic gradient methods have been shown to easily fit random labels. When there are a mixture of correct and mislabelled targets, networks tend to fit the former before the latter. This suggests using a suitable two-component mixture model as an unsupervised generative model of sample loss values during training to allow online estimation of the probability that a sample is mislabelled. Specifically, we propose a beta mixture to estimate this probability and correct the loss by relying on the network prediction (the so-called bootstrapping loss). We further adapt mixup augmentation to drive our approach a step further. Experiments on CIFAR-10/100 and TinyImageNet demonstrate a robustness to label noise that substantially outperforms recent state-of-the-art. Source code is available at https://git.io/fjsvE and Appendix at https://arxiv.org/abs/1904.11238.
Metadata
Item Type:Conference or Workshop Item (Speech)
Event Type:Conference
Refereed:Yes
Subjects:Computer Science > Artificial intelligence
Computer Science > Machine learning
DCU Faculties and Centres:DCU Faculties and Schools > Faculty of Engineering and Computing > School of Electronic Engineering
Research Initiatives and Centres > INSIGHT Centre for Data Analytics
Published in: Proceedings of the 36th International Conference on Machine Learning. International Conference on Machine Learning 97. MIR Press.
Publisher:MIR Press
Official URL:http://proceedings.mlr.press/v97/arazo19a/arazo19a...
Copyright Information:© 2019 The Authors
Use License:This item is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 3.0 License. View License
Funders:Science Foundation Ireland (SFI) under grant numbers SFI/15/SIRG/3283 and SFI/12/RC/2289
ID Code:23274
Deposited On:09 May 2019 11:58 by Diego Ortego Hernández . Last Modified 28 Apr 2022 10:27
Documents

Full text available as:

[thumbnail of 2019_ICML_final_arXiv_titleCorrected.pdf]
Preview
PDF - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
3MB
Downloads

Downloads

Downloads per month over past year

Archive Staff Only: edit this record