Ton-Mai, Kimberly Thien; (2024) Representation Learning for Anomaly Detection in Different Modalities. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text (Redacted copy) PhD_Thesis__Edited.pdf - Accepted Version Download (12MB) | Preview

Abstract

Anomaly detection is the task of identifying unusual instances that deviate from typical appearances or behaviours. Use cases include fraud detection, medicine, and fault detection. Effective automated systems can identify anomalies in situations that are too challenging or costly for humans. However, satisfactory detection performance relies on underlying representation space that depicts the training data. In this thesis, we investigate what characteristics form a good representation. We conduct experiments on images, text, speech, and tabular data to examine how well anomalies can be detected in each case and to find commonalities across the modalities. We find that no representation learning scheme performs well across all modalities. However, our results suggest that low-dimensional embeddings are best for anomaly detection. Using embeddings from pre-trained networks is an effective starting point and fine-tuning boosts performance. We also analyse how the anomaly detection architectures affect results. We show the detector is unimportant as long as the representation space is reasonable. The choice of representation should consider prior knowledge about the anomalies and how they contrast with the benign distribution. Overall, our findings suggest anomaly detection research should focus on representation learning objectives rather than modifying architectures or scoring functions.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item