Domain Shift Detection in Machine Learning Models (Master Thesis)
Machine learning models often assume that training and deployment data come from the same distribution. In practice, distribution shifts — known as domain shifts — are common and can lead to significant performance degradation. Detecting these shifts early is crucial for maintaining reliable and safe AI systems in areas such as healthcare, autonomous systems, and finance.
This project aims to explore methods for detecting and characterizing domain shifts in machine learning models. Approaches may include statistical measures, uncertainty-based techniques, and representation-based analysis to identify changes in data distributions and model behavior.
The thesis has the following objectives:
- Study existing techniques for domain shift detection, including feature-based and model-based approaches.
- Develop novel methods for identifying shifts in unlabeled or partially labeled datasets.
- Evaluate performance on benchmark datasets and real-world scenarios, analyzing robustness and sensitivity.
