[2006.07529] Rethinking the Value of Labels for Improving Class-Imbalanced Learningopen searchopen navigation menucontact arXivsubscribe to arXiv mailings

Real-world data often exhibits long-tailed distributions with heavy class imbalance, posing great challenges for deep recognition models. We identify a persisting dilemma on the value of labels in the context of imbalanced learning: on the one hand, supervision from labels typically leads to better results than its unsupervised counterparts; on the other hand, heavily imbalanced data naturally incurs "label bias" in the classifier, where the decision boundary can be drastically altered by the majority classes. In this work, we systematically investigate these two facets of labels. We demonstrate, theoretically and empirically, that class-imbalanced learning can significantly benefit in both semi-supervised and self-supervised manners. Specifically, we confirm that (1) positively, imbalanced labels are valuable: given more unlabeled data, the original labels can be leveraged with the extra data to reduce label bias in a semi-supervised manner, which greatly improves the final classifier

Date: 2020/11/22 15:51

Related Entries

Read more INRIA ALMANACH - Transfer Learning - Google Slides
0 users, 1 mentions 2020/03/28 12:51
Read more [2002.03049] Snippext: Semi-supervised Opinion Mining with Augmented Dataopen searchopen navigation ...
0 users, 1 mentions 2020/06/16 14:21
Read more BERT — A Practitioner’s Perspective | by Nirupam Purushothama | The Startup | Jul, 2020 | Medium
0 users, 1 mentions 2020/07/31 09:30
Read more Permutation Equivariant Models for Compositional Generalization in Language | OpenReview
0 users, 1 mentions 2020/08/18 15:51
Read more [2004.13625] Event Extraction by Answering (Almost) Natural Questionsopen searchopen navigation menu...
0 users, 1 mentions 2020/10/12 14:22
Read more [2010.03274] Learning to Explain: Datasets and Models for Identifying Valid Reasoning Chains in Mult...
0 users, 1 mentions 2020/10/27 14:21