M.S. Candidate: Özge Köktürk
Program: Data Informatics
Date: 06.01.2025 / 14:30
Place: A-212

Abstract: In practical use of machine learning models, generalizability is of crucial importance. When a model is trained on a dataset obtained in a specific context, it often performs poorly in similar situations but under different contexts. This can lead to unreliable predictions and potentially harmful decisions in real-life applications. This thesis proposes a training methodology for context-invariant autoencoders through unsupervised domain adaptation, aiming to learn representations that remain stable across varying contexts. Consequently, any application can be built on top of these domain-invariant representations. In this study, domain-adversarial training and data augmentation strategies have been employed to extract features that capture the essential structures of input images while disregarding features associated with contextual changes. For the experiments, image data collected from the CARLA (Car Learning to Act) simulator system under different weather conditions and various times of day have been used.