Abstract

One of the biggest challenges of employing supervised deep learning approaches is their inability to perform as well beyond standardized datasets in real-world applications. Therefore, abrupt changes in the form of an outlier or overall changes in data distribution after model deployment result in a performance drop. Owing to these changes that induce distributional shifts, we propose two methodologies; the first is the detection of these shifts, and the second is adapting the model to overcome the low predictive performance due to these shifts.
The former usually refers to anomaly detection, the process of finding patterns in the data that do not resemble the expected behavior.
Understanding the behavior of data by capturing their distribution might help us to find those rare and uncommon samples without the need for annotated data. In this thesis, we exploit the ability of generative adversarial networks (GANs) in capturing the latent representation to design a model that differentiates the expected behavior from deviated samples. Furthermore, we integrate self-supervision into generative adversarial networks to improve the predictive performance of our proposed anomaly detection model. In addition, to shift detection, we propose an ensemble approach to adapt a model under varied distributional shifts using domain adaptation.
In summary, this thesis focuses on detecting shifts under the umbrella of anomaly detection as well as mitigating the effect of several distributional shifts by adapting deep learning models using a Bayesian and information theory approach.