Locatello Group

Causal Learning and Artificial Intelligence

Modern machine learning is still limited to a superficial description of reality that only holds when the experimental conditions are fixed. It largely ignores interventions in the world, domain shifts, and temporal structure. Locatello’s group focuses on learning causal representations and causal models from data to address these issues. Through rigorous theory and scalable algorithms, they enable AI agents to understand cause-and-effect relationships, the effect of interventions, and distribution changes underlying the data they encounter.

Advances in causal learning have promising applications in machine learning and artificial intelligence, including robustness, explainability, and fairness in recognition, reasoning, and planning tasks. Most importantly, their impact extends to discovering new scientific knowledge from massive amounts of data and serving as the interface between people and complex systems for decision-making.

Causal learning emerges by discovering structural knowledge about the data-generating process and meaningful abstractions that allow for well-defined causal relations. Research questions span theoretical identifiability and methodological innovations in causal learning, discovering objects and abstract entities, their factors of variations, and their causal relations. The group applies these methods to solve open problems in machine learning and enable new applications in the sciences.