KDL studies how to construct causal models of complex systems, a fundamental research challenge at the frontier of machine learning. In particular, we create new methods, algorithms, and systems that infer causal dependence from observational and experimental data about complex and time-varying relationships among people, places, things, and events. Current research focuses on several areas, including: (1) using causal models to provide human-understandable explanations of how deep neural networks make inferences; (2) using causal models to assess the competence of machine learning models (the circumstances under which the models will perform well or poorly); (3) learning causal models that provide accurate inferences when presented with novel inputs; and (4) methods for effective evaluation of methods for causal modeling.
New developments in causal inference are vital because of growing interest in moving beyond simple predictive models, toward models that can correctly infer the effects of actions. Such models are critical to designing, managing, and understanding AI systems, the internet, cyber-physical systems, scientific communities, financial systems, social networks, complex software, and other types of complex systems.
Our research draws on concepts and techniques from a wide variety of technical communities, including machine learning, graphical models, probabilistic programming, statistics, experimental and quasi-experimental design, quantitative social science, database theory, complex adaptive systems, graph theory, and social network analysis. Our work intentionally spans the spectrum from foundational theory of statistical inference to large-scale empirical evaluation of the resulting algorithms and systems.