Theory-Based Bayesian Models of Inductive Inference.

摘要

Our research aims to develop computational models of inductive inference in higher-level human cognition, specifically the ability to generalize from sparse, noisy, ambiguous data, and to build more human-like machine learning systems. Work has focused on two areas of cognition: learning about categories and their properties, and learning about relational structures -- specifically, systems of causal and social relations. We have developed a theory-based Bayesian framework for modeling these learning tasks as statistical inference over hierarchies of structured knowledge representations. Our models have made several contributions. First, they have explained a broad range of phenomena with high quantitative accuracy, using a minimum of free parameters. Second, they have provided a rational framework for explaining how and why everyday induction works, in terms of approximations to optimal statistical inference in natural environments. Third, they have provided tools for elucidating people's implicit theories about the structure of the world - describing the form and content of the prior knowledge that guides inductive inference and explaining how it may be acquired from experience. Finally, our models have led to improved algorithms for machine learning of category structures and their property distributions, causal networks, and the structure of social relations, thus bringing artificial intelligence systems closer to the capacities of human intelligence at the same tone as they support a better understanding of human intelligence.