Class-Imbalanced Deep Learning via a Class-Balanced Ensemble.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Class imbalance is a common issue in real-world datasets, posing challenges for machine learning models, especially deep learning.
• Traditional methods often struggle to effectively learn from imbalanced data, leading to biased models that favor majority classes.

Purpose of the Study:

• To develop a novel method for tackling class-imbalanced learning problems in deep convolutional neural networks (CNNs).
• To enhance the performance of CNNs in scenarios where data distribution is uneven across classes.

Main Methods:

• Embedding an ensemble of auxiliary classifiers into CNNs, trained end-to-end with the main network.
• Designing a new loss function that guides hidden layers and auxiliary classifiers to focus on misclassified samples, promoting diverse learning.
• Utilizing auxiliary classifiers to assist the main CNN during training or acting as a standalone component post-training.

Main Results:

• Demonstrated significant performance improvements on four benchmark datasets (CIFAR-10, CIFAR-100, iNaturalist, CelebA) with increasing complexity.
• Outperformed existing state-of-the-art deep learning methods for imbalanced data.
• The proposed ensemble approach effectively mitigates bias towards majority classes.

Conclusions:

• The proposed ensemble learning method integrated with CNNs is effective in addressing class imbalance.
• The auxiliary classifiers enhance the CNN's ability to learn from imbalanced data, leading to improved classification accuracy.
• This approach offers flexibility, allowing auxiliary classifiers to aid training or be removed post-training.