Enhancing out-of-distribution detection with bilateral distribution score.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Out-of-distribution (OOD) detection is vital for trustworthy AI in safety-critical systems.
• Existing methods struggle with classifier overconfidence and deceptive OOD samples.
• Current post-hoc OOD detection methods lack robust performance against certain OOD data distributions.

Purpose of the Study:

• To develop a novel OOD detection method that overcomes limitations of existing approaches.
• To introduce a new OOD scoring mechanism based on an "ideal OOD sample" concept.
• To enhance the reliability and safety of machine learning models in real-world applications.

Main Methods:

• Proposed the concept of an "ideal OOD sample" equidistant from all class centers in feature space.
• Defined a novel OOD score based on similarity to this ideal OOD sample.
• Introduced the Bilateral Distribution Score (BDS) integrating both OOD and in-distribution (ID) scores.

Main Results:

• BDS demonstrated superior OOD detection capabilities on ImageNet-1k and CIFAR-10 benchmarks.
• Achieved a 10.78% reduction in average False Positive Rate at 95% (FPR95) compared to state-of-the-art methods.
• The method requires no architectural modifications or retraining, ensuring backward compatibility.

Conclusions:

• BDS offers a significant advancement in OOD detection, improving model trustworthiness.
• The proposed method is effective and practical, integrating seamlessly with existing techniques.
• BDS enhances the robustness of machine learning models against OOD samples.