Learning Representations for Novelty and Anomaly Detection
Ranya Almohsen · 2023
The problem of novelty or anomaly detection refers to the ability to automatically identify data samples that differ from a notion of normality. Techniques that address this problem are necessary in many applications, like in medical diagnosis, autonomous driving, fraud detection, or cyber-attack detection, just to mention a few. The problem is inherently challenging because of the openness of the space of distributions that characterize novelty or outlier data points. This is often matched with the inability to adequately represent such distributions due to the lack of representative data. In this dissertation we address the challenge above by making several contributions. (a)We introduce an unsupervised framework for novelty detection, which is based on deep learning techniques, and which does not require labeled data representing the distribution of outliers. (b) The framework is general and based on first principles by detecting anomalies via computing their probabilities according to the distribution representing normality. (c) The framework can handle high-dimensional data such as images, by performing a non-linear dimensionality reduction of the input space into an isometric lower-dimensional space, leading to a computationally efficient method. (d) The framework is guarded from the potential inclusion of distributions of outliers into the distribution of normality by favoring that only inlier data can be well represented by the model. (e) The methods are evaluated extensively on multiple computer vision benchmark datasets, where it is shown that they compare favorably with the state of the art.