Public defence in Computer Science, M.Sc. Çağlar Hizli

Title of the thesis: Causality, identifiability, and representation learning for machine learning with non-i.i.d. data

Thesis defender: Çağlar Hizli 
Opponent: Assistant Professor Francesco Locatello, Institute of Science and Technology, Austria
Custos: Associate Professor Pekka Marttinen, Aalto University School of Science

Modern AI systems perform impressively when training data and test data are similar. Real life is different: patients, environments and tasks change over time, and data are often limited or expensive to collect. This doctoral thesis examines how machine learning could become more reliable in such changing conditions.

The thesis focuses on learning from and generalizing to data that do not follow the usual assumption of being stable and uniform. Its aim is to help build AI systems that can handle changing environments, learn from small amounts of data, and transfer what they have learned to new tasks.

The work addresses this challenge from three complementary directions. First, in healthcare time-series data, it shows how to estimate what would happen to patients if treatment policies changed. In studies of meal and blood-glucose data, the methods produced clinically meaningful predictions. In obesity patients undergoing bariatric surgery, they also helped reveal how surgery changes glucose profiles through meal intake and other metabolic pathways.

Second, the thesis introduces a new theoretical framework for recovering hidden dynamical processes from complex observations collected in changing environments. It shows that identifying these stable underlying processes improves the ability to predict future progression.

Third, in computer vision, the thesis proposes an object-centred training method that learns visual representations which transfer better to new tasks and datasets. The method improved results not only on the standard ImageNet benchmark but also on thirteen datasets designed to test performance under changed conditions.

The thesis is relevant to a broad range of fields, including healthcare, scientific modelling and other AI applications where conditions change over time. The central conclusion is that future intelligent systems should combine lower-level pattern-recognition skills with higher-level cognitive abilities, such as understanding cause and effect, recovering stable mechanisms across changing environments, and representing the world in terms of objects and their relations.

Keywords: causality, identifiability, time-series, object-centered learning

Thesis available for public display 7 days prior to the defence at Aalto University's public display page.