E ³Outlier: a Self-Supervised Framework for Unsupervised Deep Outlier Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised outlier detection (OD) struggles with large visual datasets.
• Deep neural networks (DNNs) are effective for visual data but challenging for unsupervised OD.

Purpose of the Study:

• Propose E³Outlier, a novel framework for effective, end-to-end deep outlier removal.
• Introduce self-supervision into deep OD to address limitations of existing methods.

Main Methods:

• Employ a discriminative learning paradigm with pseudo-classes from unlabeled data.
• Utilize inlier priority during self-supervised learning for end-to-end OD.
• Explore network uncertainty as an outlierness measure with score refinement strategies.

Main Results:

• E³Outlier significantly outperforms state-of-the-art methods, achieving 10%-30% higher AUROC.
• Demonstrated effectiveness across various datasets and potential for video abnormal event detection.
• Explored generative and contrastive learning paradigms for further performance gains.

Conclusions:

• E³Outlier provides an effective and extendable solution for unsupervised deep outlier detection.
• Self-supervision, inlier priority, and network uncertainty are key to successful deep OD.
• The framework shows promise for diverse OD applications beyond image analysis.