Amir Mohammad Karimi Mamaghan

Causality and Representation Learning are foundational to advancing AI systems capable of reasoning, generalization, and understanding the complex structure of the world. Causality provides tools to uncover the underlying structure of a system, understand cause-effect relationships, and reason about interventions. Representation learning, on the other hand, transforms raw data into structured abstractions that are essential for modeling systems and supporting decision-making. Both paradigms aim to lay the groundwork for models that are both interpretable and generalizable.

This thesis is structured in two main parts. In the first part, we conduct a principled analysis of current evaluation protocols in bayesian causal discovery and object-centric learning. The goal is to understand where each family of methods excels, where their evaluation protocols fall short, and what truly matters for downstream reasoning. Building on these insights, the second part of the thesis explores how object-centric inductive biases can be extended to part-whole hierarchies, capturing not just entire objects, but also their constituent parts in a self-consistent and controllable way. We investigate how to leverage foundation models alongside object-centric inductive biases to model part-whole hierarchies and evaluate them on tasks that require multi-level reasoning. By identifying when object-centric models pay off and how they can be scaled or combined, this thesis contributes to the development of vision systems that parse the world into more meaningful and human-interpretable components.