Anomaly detection via Gumbel Noise Score Matching

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Anomaly detection is crucial for identifying unusual patterns in data.
• Existing methods may struggle with categorical data or specific applications like image segmentation.
• Unsupervised learning offers a way to detect anomalies without labeled data.

Purpose of the Study:

• To introduce Gumbel Noise Score Matching (GNSM), a novel unsupervised method for anomaly detection.
• To evaluate GNSM's performance on various tabular datasets.
• To demonstrate GNSM's applicability to image data, specifically for detecting poor segmentation predictions.

Main Methods:

• Gumbel Noise Score Matching (GNSM) estimates scores of continuously relaxed categorical distributions.
• The method utilizes a score matching training objective.
• An open-source implementation is provided.

Main Results:

• GNSM achieved consistently high performance across multiple tabular anomaly detection datasets.
• Applied to image segmentation, GNSM successfully identified images with segmentation failures.
• Anomaly scores from GNSM strongly correlated with segmentation metrics against ground truth.

Conclusions:

• GNSM is a flexible and effective unsupervised anomaly detection technique.
• The method shows promise for both tabular data and complex tasks like image segmentation quality assessment.
• GNSM offers a valuable tool for identifying anomalies in diverse data types.