Simon Bing

Causal models offer great promise to better describe and understand complex systems. While such methods have been successfully applied in settings where the variables of interest are observed, their application to unstructured, high-dimensional data remains ellusive. Causal representation learning aims to unite the capability of modern machine learning methods to process such high-dimensional data with the principled mechanistic modelling approach of causal inference. We focus on problem settings where spatio-temporal data arrises, such as climate science or neuroscience, and aim to exploit this structure in the data. Since causal representation learning is heavily underconstrained, we explore how assumptions can be derived from explicit downstream causal queries, in order to develop pragmatic algortihms that yield representations with the necessary characteristics to answer the queries of interest. This bottom-up approach leads us to reconsider the classical problem formulation of causal representation learning, yielding new understandings of how machine learning and causal inference can benefit from cross-pollination.