M.Sc. Shayan Gharib defends his PhD thesis “Learning under Data Shift: From Distributional Invariance to Structural Adaptation” on Friday the 11th of September 2026 at 13 in the University of Helsinki Main Building, Hall F3017 (Fabianinkatu 33, 3rd floor). His opponent is Associate Professor Jenni Raitoharju (University of Jyväskylä) and custos Professor Arto Klami (University of Helsinki). The defence will be held in English.
The thesis of Shayan Gharib is a part of research done in the Department of Computer Science and in the Multi-source Probabilistic Inference group at the University of Helsinki. His supervisor has been Professor Arto Klami (University of Helsinki).
Learning under Data Shift: From Distributional Invariance to Structural Adaptation
Machine learning has established itself as a pivotal technology, empowering significant progress in fields as diverse as industrial automation and autonomous navigation. However, the success of these systems typically hinges on a "static world" assumption: the belief that the conditions encountered during deployment will perfectly mirror those seen during training. In reality, the world is dynamic. Changes in environment, sensor aging, or even the subjective biases of human annotators create a phenomenon known as data shift, which can cause once-reliable models to fail catastrophically. This thesis explores the fundamental challenge of ensuring that machine learning remains adaptable when faced with such inevitable changes.
We frame this challenge as a progression from distributional invariance to structural adaptation. In computer vision tasks, we address covariate shift, where changes such as variations in lighting or background alter the appearance of objects without changing their identity. Such variations often lead to performance degradation as the model's learned patterns fail to generalize to new observations. To address this, we propose strategies to compensate for the performance drop by learning representations that are invariant to these environmental fluctuations, effectively teaching models to focus on the essence of an image rather than its environmental artifacts. We then extend this perspective to the ``noisy'' world of weak supervision, where we interpret the challenge of learning from incomplete labels as a form of concept shift. By leveraging structural information in the form of label cardinality, we demonstrate how to reconstruct reliable models from biased and limited observations.
The final part of this thesis moves beyond deep-learning based methods to address the physical reality of sensing networks. In structural health monitoring, data shifts are often caused by tangible changes, such as the relocation of a transducer or the gradual accumulation of rust and fouling. For these scenarios, we introduce a self-adaptation framework that utilizes a physics-informed forward simulation to align a model with its altered environment.
In summary, this thesis provides a unified perspective on managing diverse data shifts across a range of practical applications. The proposed strategies demonstrate that successful adaptation is achievable across a spectrum ranging from static environmental shifts in image data to dynamic structural changes in physical sensing networks by actively adapting to changes and ensuring reliable operation in ever-evolving conditions.
Availability of the dissertation
An electronic version of the doctoral dissertation will be available in the University of Helsinki open repository Helda at
Printed copies will be available on request from Shayan Gharib: