Mitigating the Adverse Effects of Long-Tailed Data on Deep Learning Models
Communications in Computer and Information Science
When the data distribution in a dataset is highly imbalanced or long-tailed, it can severely affect the effectiveness of a deep network model. This drop in performance is caused due to the biased classifier, which favours the head-class samples because these samples have more dominant features compared to the tail-class samples. Addressing this challenge requires not only capturing subtle inter-class differences and intra-class similarities but also effectively utilising limited data for the minority classes. Supervised contrastive learning (SCL) and transfer of angle information from head classes to tail classes have recently been proposed to address the problem of long-tail classification. For a well-balanced dataset, SCL demonstrates effectiveness by pulling together samples from the same classes while pushing away samples from different classes. However, when applied to long-tailed datasets, SCL could become biased towards the head-class samples. On the other hand, the method of transfer of angle information aims to address the challenges posed by long-tailed image classification; however, it lacks in achieving both intra-class compactness and inter-class separability. To address the shortcomings and exploit the strengths of both of these approaches, we propose a unique hybrid method that seamlessly integrates supervised contrastive learning and angular variance to mitigate the adverse effects of long-tailed data on deep learning models for image classification. We name our method as Supervised Angular Contrastive Learning (SACL). In our experiments on long-tailed datasets with different class imbalance ratios, we demonstrate that our method outperforms most of the existing baseline approaches.
Open Access Status
This publication is not available as open access