Danru Xu

Endowing machine learning models with causal reasoning capabilities is a promising direction for improving their generalization, robustness, and interpretability. Current machine learning models have shown impressive performance in many tasks, but are still often limited to learning correlations, which can make them brittle to distribution shifts and prone to bias, hindering their general applicability. Conversely, most of causality research focuses on estimating the effects of actions and identifying causal relationships from data, but assumes that we have handcrafted features or well-defined causal variables that we are interested in.

Causal representation learning (CRL) is a promising new direction in causality that combines the strengths of deep learning, especially representation learning, with the theoretical guarantees of causality. In particular, CRL focuses on identifying causal variables and causal relations between them from high-dimensional and unstructured data, e.g. images, video or text, that represent a measurement of an underlying causal system. While one cannot in general fully recover the ground truth causal variables and their mechanisms from unstructured data in an unsupervised way, CRL methods propose different sets of assumptions under which one can identify the causal variables up to an equivalence class. Typically, the goal is to learn a disentangled representation of the causal variables, even if they are causally related, and hence not independent, as in typical disentanglement research.

While CRL is an exciting new field with a lot of momentum, most current works consider a single modality and assume that all causal variables are captured in the high-dimensional observations. This PhD project focuses on learning causal representations in complex settings, e.g. when we have multimodal data, partial observations from multiple sources, or partial observability in an environment. Additionally, it aims at integrating existing background knowledge, e.g. known causal relations or known hierarchies, to improve the learning and generalization of the learned causal representations, a research direction that has been completely unexplored until now. Finally, it proposes to develop new methods to generalize and adapt causal representations to new unseen environments. Through these goals, this PhD project aims at showing the potential of combining causality and representation learning to enhance the robustness, generalization, and interpretability of future machine learning models.