Funded Grants

Compositionality in probabilistic models of cognition

Structured mental representations allow humans to impose order on the tangle of experience. I study the computational basis of mental representations that underly higher-level cognition by using a combination of formal modeling and behavioral experimentation. The ultimate goal of my work is to develop a formal theory of the structure of thought that accounts for the systematic and flexible inferences central to human cognition in the real world.

To look for broad computational principles, it is useful to consider a broad range of cognitive abilities. My work draws case studies from causal reasoning, category learning, and social cognition. Across these areas I aim to understand the primitive elements of mental representation, the ways they combine into complex thoughts and stable representations (concepts), and how these thoughts enable us to respond successfully to the demands of our environment.

This is a difficult project, but computational modeling provides formal tools to cut nature at its joints: mathematically precise psychological theories, and frameworks for reasoning out their consequences. My approach integrates probabilistic inference techniques with compositional representations drawn from logic and semantics--together these give systems with the systematic structure of a language and the robust reasoning under uncertainty of statistical induction. These theoretical ideas have crystallized into the probabilistic language of thought hypothesis--a collection of informal principles of mental representation, together with a precise mathematical realization of these principles.

If formal modeling is the warp of my work then empirical observation is the weft. I use a variety of experimental techniques to test and refine models. These include: novel behavioral paradigms with adult participants, for instance to study grounded causal learning; collaborations with developmental psychologists to study pre-schoolers and infants; and, increasingly, "high throughput" data collection using the internet. In future work I plan to incorporate other methodologies, such as cross-cultural behavioral experiments.

In the coming years I will continue theoretical and empirical studies exploring the probabilistic language of thought hypothesis, emphasizing three key themes that have emerged in my recent work. First, that mental representations gain their meaning, and usefulness, partly through their grounding in concrete perception and action. Second, that real-world cognition relies on the interaction between different domains of knowledge--e.g. joint reasoning about causes, objects, and agents. Third, that abstract mental representations arise developmentally from a powerful inductive learning mechanism aided by specialized perceptual supports and cultural input. My metric of success in these studies will be the ability to predict and explain experimental studies, the ability to use the principles discovered to engineer intelligent systems, and the contribution of these principles to a broad understanding of human cognition.