Dynamically balancing class losses in imbalanced deep learning. Issue 5 (29th December 2021)

• Record Type: Journal Article
• Authors: Zhao, Yaochi; Liu, Shiguang; Hu, Zhuhua
• Is Part Of: Electronics letters. Volume 58:Issue 5(2022)
• Page Start: 203
• Page End: 206
• Publication Date: 2021-12-29
• Subjects: Electronics -- Periodicals; 621.381
• DOI: 10.1049/ell2.12408
• Languages: English
• ISSNs: 0013-5194

Abstract

Abstract: Imbalanced datasets commonly exist in real world and greatly challenge the performances of deep neural models. The authors discover that the traditional balance strategies in the existing imbalanced learning methods emphasize/suppress the importance of minority/majority class in a fixed way. Even the minority class is fully represented, the minority/majority class is still emphasized/suppressed, eventually leading to sub‐optimal models. The authors argue that the importance of minority/majority class should be emphasized/suppressed in a dynamic manner. They propose to dynamically balance class losses (DBCL), which first suppresses the importance of majority class, and then gradually restores the importance of majority class. By adding the idea of DBCL into three state‐of‐the‐art cost‐sensitive loss functions, the authors propose three novel loss functions: dynamic constrained‐focal‐loss (DCFL), dynamic balance‐cross‐entropy based on the effective number of samples (DBENS) and dynamic balance‐cross‐entropy based on class frequencies (DBCEF). They apply DCFL, DBENS and DBCEF to the imbalanced learning in salient‐instance segmentation. The results show that their methods significantly improve the classification abilities of models.