Novel Regularization via Logit Weight Repulsion for Long-Tailed Classification
Taegil Ha (Seoul National University),* Seulki Park (Seoul National University), Jin Young Choi (Seoul National University)
The 34th British Machine Vision Conference
Abstract
Research to address the class imbalance problem aims to balance the impact of each class on the loss function because logit weight vectors tend to favor a majority class. To this end, researchers have introduced balanced losses such as margin-based loss and logit adjustment. The balanced losses succeed to classify the minority class better than the conventional loss. However, the balanced loss focuses on balancing the norm of logit weight, but overlooks the direction of logit weight vectors. As a result, the balanced loss sacrifices the head class performance by shrinking the region between the logit vectors. In this paper, we delve into the behavior of the gradient of the balanced loss and clarify how it shrinks the decision plane of each class from two perspectives. First, balanced loss pushes the decision boundary from tail to head within limited space, shrinking the decision plane of the head class. Second, balanced loss does not prevent the logit vectors to have a similar direction to each other during the update, shrinking region between logit vectors. Based on this study, we propose a new regularization called Logit Weight Repulsion (LWR), which encourages a logit weight vector for a class to repel those for other classes. This repulsion enlarges the region between the logit vectors for each class. The proposed LWR regularizer has been evaluated on benchmark datasets where ours achieves the state-of-the-art performance for long-tailed classification. Notably, LWR achieves performance improvements in minority classes without sacrificing the performance in majority classes.Video
Citation
@inproceedings{Ha_2023_BMVC,
author = {Taegil Ha and Seulki Park and Jin Young Choi},
title = {Novel Regularization via Logit Weight Repulsion for Long-Tailed Classification},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0762.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by