Towards out-of-distribution detection using gradient vectors.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Real-world deep learning deployment necessitates handling uncontrolled input data.
• Out-of-distribution (OOD) detection is crucial for identifying samples outside known classes.
• Existing OOD methods often overlook the potential of gradient space.

Purpose of the Study:

• To introduce GradVec, a new family of OOD detection methods.
• To explore the utility of gradient space for OOD detection.
• To improve the robustness of deep learning models in real-world scenarios.

Main Methods:

• Proposed GradVec methods leverage gradient features for OOD detection.
• Utilized gradient space as an input representation for OOD detection techniques.
• Applied GradVec to pre-trained models without altering the training procedure.

Main Results:

• GradVec demonstrated superior performance in OOD detection tasks.
• Achieved significant reductions in False Positive Rate at 95% FPR (FPR95) for image classification (up to 26.67%).
• Showcased effectiveness in text classification, reducing FPR95 by up to 21.29%.

Conclusions:

• Gradient space offers an informative representation for OOD detection.
• GradVec provides a versatile and effective approach for enhancing model safety.
• The method is applicable to any pre-trained model without additional data or training modifications.