When an extra rejection class meets out-of-distribution detection in long-tailed image classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Reliable deep learning requires effective Out-of-Distribution (OOD) input detection.
• Existing OOD detection methods struggle with imbalanced, long-tailed training data distributions.
• Open-world applications necessitate robust OOD detection for dependable AI systems.

Purpose of the Study:

• To develop an effective OOD detection framework for long-tailed image classification.
• To address the limitations of current OOD detection methods on imbalanced datasets.
• To improve the reliability of deep learning models in real-world, open-world scenarios.

Main Methods:

• A three-branch training framework incorporating a rejection class and auxiliary outlier data.
• Assigning all outlier training samples to the rejection class label.
• Utilizing an inlier loss, outlier loss, and Tail-class prototype induced Supervised Contrastive Loss (TSCL) for unified network training.

Main Results:

• The proposed method demonstrates superior OOD detection performance in long-tailed image classification.
• Achieved a 1.23% improvement in average AUROC and a 3.18% reduction in average FPR95 on CIFAR100-LT compared to Outlier Exposure (OE).
• The OOD detector constructed using the rejection class proved effective during inference.

Conclusions:

• The proposed three-branch framework effectively enhances OOD detection for long-tailed distributions.
• Incorporating a rejection class and auxiliary outlier data is crucial for robust OOD detection.
• The method offers a significant improvement over existing techniques for real-world deep learning applications.