Anomaly detection with deep learning methods

Anomaly detection is the process of identifying data samples that deviate from the normal and has significant applications in various fields. This thesis investigates deep learning methods for anomaly detection and proposes a new method using deep hypersphere classifiers. In the proposed method, unlike other deep hypersphere classifiers, the hypersphere center is considered as a learnable parameter and this center is updated according to the changing deep feature representations during training. Additionally, a more robust loss term against noisy labels in the datasets used for Outlier Exposure has been proposed. The experiments conducted have shown that the proposed method achieves the best or equivalent results in the literature on commonly used datasets for anomaly detection.