Structure and Causality in Understanding Complex Systems

Abstract

A central goal of science and engineering is to understand the causal structure of complex computational, physical, and social systems. Inferring this causal structure without performing experiments, however, is often extremely challenging. This thesis develops new mathematical approaches for exploiting the structure underlying many types of data to reveal insights about the causal relationships governing complex systems. The work consists of four aims, each of which leverages structure and causal modeling to understand a different type of system. In the first aim, we develop an algorithm based on the sparse Bayesian learning (SBL) framework for exploiting sparse and temporal structure in order to more efficiently collect data from time-varying high-dimensional systems. In the second aim, we develop a framework for explaining the operation of black-box machine learning classifiers using a causal model of how the data and classifier output are generated. In the third aim, we analyze a class of algorithms that use low-dimensional structure to infer causal interactions in coupled dynamical systems. In the final aim, we use surveys of the public and AI practitioners to model attitudes toward artificial intelligence adoption and governance, and employ the model to answer policy-relevant questions about AI governance.