VISAL-A novel learning strategy to address class imbalance

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Deep Neural Networks (DNNs) exhibit poor generalization on minority classes in imbalanced datasets.
• Addressing data imbalance is crucial for reliable classification performance.

Purpose of the Study:

• To introduce Visually Interpretable Space Adjustment Learning (VISAL), a novel learning function for imbalanced data classification.
• To enhance the generalization capability of minority class samples within DNNs.

Main Methods:

• VISAL integrates angular and Euclidean margins into the cross-entropy learning strategy.
• The method aims to create additional space for minority class samples to improve model generalization.

Main Results:

• VISAL was evaluated on imbalanced versions of CIFAR, Tiny ImageNet, COVIDx, and IMDB reviews datasets.
• The proposed VISAL method significantly outperformed existing state-of-the-art approaches.

Conclusions:

• VISAL offers a simple yet effective solution for imbalanced data classification tasks.
• The technique demonstrates superior performance in improving minority class generalization for Deep Neural Networks.