Dario Rancati

Understanding Data Generating Processes is one of machine learning's fundamental tasks, both from the foundational standpoint and with astounding empirical applications in fields such as language modelling, computer vision and decision theory. Deep Generative Models are a vast class of methods to model DGPs encompassing techniques such as GANs, VAEs, Normalizing Flows and Diffusion Models. While they possess state-of-the-art performances on numerous tasks, they are not without shortcomings: most notably, the latent space these models rely on often lacks any tangible form of interpretability, leading to difficulties generalizing out-of-domain and learning disentangled representations. Conversely, the field of causality offers powerful tools to model the cause-effect relationships between statistical variables, defining ideas such as interventions on a system and counterfactuals (which are at the core of the field of ML interpretability) in a rigorous way. A cornerstone model in the theory of causality are structural causal models (SCMs), which represent the cause-effect relationships via edges in a DAG where the vertices are the random variables analyzed. Nevertheless, these models have their own set of limitations: they typically struggle in high-dimensional settings, and for this reason have so far had limited impact in fields such as computer vision. A clear opportunity for synergy arises between these two fields: DGMs can help causality tackle these high-dimensional settings where it typically struggles, and causality can in turn provide DGMs with interpretable capabilities they can leverage to introduce key concepts such as interventions and counterfactuals, as well as improving out-of-distribution sampling abilities.