Dynamically Weighted Balanced Loss: Class Imbalanced Learning and Confidence Calibration of Deep Neural Networks

• Published in: IEEE transaction on neural networks and learning systems Vol. 33; no. 7; pp. 1 - 12
• Main Authors: Fernando, K. Ruwani M., Tsokos, Chris P.
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.07.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Artificial neural networks; Boosting; Calibration; Classification; Confidence calibration; convolutional neural networks (CNNs); cost-sensitive learning; data imbalance; Datasets; Deep learning; Heuristic algorithms; Learning systems; loss functions; Machine learning; Medical imaging; Neural networks; Neurons; Statistical analysis; Task analysis; Training
• ISSN: 2162-237X; 2162-2388; 2162-2388
• DOI: 10.1109/TNNLS.2020.3047335

Imbalanced class distribution is an inherent problem in many real-world classification tasks where the minority class is the class of interest. Many conventional statistical and machine learning classification algorithms are subject to frequency bias, and learning discriminating boundaries between the minority and majority classes could be challenging. To address the class distribution imbalance in deep learning, we propose a class rebalancing strategy based on a class-balanced dynamically weighted loss function where weights are assigned based on the class frequency and predicted probability of ground-truth class. The ability of dynamic weighting scheme to self-adapt its weights depending on the prediction scores allows the model to adjust for instances with varying levels of difficulty resulting in gradient updates driven by hard minority class samples. We further show that the proposed loss function is classification calibrated. Experiments conducted on highly imbalanced data across different applications of cyber intrusion detection (CICIDS2017 data set) and medical imaging (ISIC2019 data set) show robust generalization. Theoretical results supported by superior empirical performance provide justification for the validity of the proposed dynamically weighted balanced (DWB) loss function.