Fahimeh Hosseininoohdani

Recent advances in large-scale deep learning models have demonstrated substantial improvements across a wide range of tasks. Nevertheless, their capacity for out-of-distribution generalization and the interpretability of their learned representations remain limited and poorly understood. A fundamental component of generalization lies in a model's capacity to decompose inputs into meaningful components and to generalize to novel combinations of the learned components. Disentangled representation learning and object-centric learning are two promising research directions that aim to approximate the inverse of the data-generating process by inducing structured representations that correspond to the underlying factors of variation in the inputs. Such representations are expected to inherently enhance models' capacity for compositional generalization. However, a principled theoretical understanding of the conditions under which these representations guarantee provable compositional generalization has only recently begun to emerge.

This project aims to develop methods for disentangled and object-centric representation learning with theoretical identifiability guarantees, and to investigate how such representations can improve models' extrapolation, compositional generalization, and robustness to distributional shifts, while providing theoretical justifications for these improvements under specific, well-defined assumptions about the data and model architecture.