Oskar Åström presents his licentiate work "Real-World Applications of Anomaly Detection"

Opponent: Aleksis Pirinen

Examiner: Johan Lindström

Chairman: Magnus Oskarsson

Abstract. This thesis studies anomaly detection in real-world settings, where tasks are often ill-posed and defined by identifying unexpected behaviour rather than predicting fixed targets. Because complex high-dimensional data such as images are difficult to model with traditional parametric distributions, the thesis focuses on neural probabilistic models that learn data distributions from examples, and evaluates their practical usefulness on downstream applications.

The thesis makes three contributions. First, it develops a method for modelling conditional distributions in variational autoencoders (VAEs) and shows that introducing slight rigidity in class-cluster positions can improve generalization. Second, it applies this approach to point-of-care ultrasound (POCUS) for breast cancer assessment by modelling the distribution of correctly acquired images to detect and filter scans with artifacts that hinder diagnosis. Third, it models spatial crop-yield distributions using graph neural networks and Sentinel-1/2 remote sensing data to estimate historical yield patterns and forecast future variability, supporting more informed fertilization decisions and reducing environmental impact.

The thesis is available in the Mathematical Library and in LUCRIS through the following link.

https://portal.research.lu.se/en/publications/real-world-applications-of-anomaly-detection-detecting-the-unexpe/