Speaker: Jorge Medina Moreira
<u>EDIC candidacy exam</u><br>
Exam president: Prof. Amir Zamir<br>
Thesis advisor: Prof. Lenka Zdeborová<br>
Co-examiner: Prof. Michael Gastpar<br>
<br>
<u>Abstract</u><br>
Self-supervised learning has become a central paradigm for training modern machine learning models, yet its theoretical foundations remain incomplete. This thesis studies self-supervised feature learning through the lens of statistical physics, random matrix theory, and high-dimensional inference. Focusing on three recent theoretical works, it examines how different self-supervised objectives induce implicit biases i