Computational Principles of Intelligence
We are an interdisciplinary research group that develops computational models of human intelligence. Our goal is to build formal theories of how people generalize from little data, explore efficiently, and find approximate solutions to complex problems.
We build precise and powerful models of people's cognitive abilities, combining psychology, machine learning, and neuroscience. We use interactive games, large data sets, online and lab experiments to study how people learn and explore. We focus on the following three topics:
Compared to machine learning algorithms, people are generally much better at generalizing from limited data. To account for this, we work on compositional theories of generalization. Our account assumes people rely on compositional inductive biases: priors over different structural forms that can be combined and reused, creating potentially infinite generalizations from a finite set of simple building blocks. We model human generalization using methods of function approximation and program induction.
We study how people use structure to guide their search for rewards. Our models combine the ability to generalize with an uncertainty-driven exploration strategy. These models describe how adults use correlational structure to find rewards, how children and adults differ in their search behavior, and how people explore complex naturalistic environments. However, a simple drive towards uncertainty might not be all there is to human curiosity. We therefore investigate alternative forms of exploration based on an learner's competence or different information measures.
We investigate how people trade off between accuracy and efficiency when solving complex problems. As biological computation normally costs time and energy, a computationally efficient agent might halt computation after a short time. Using the notion of computational rationality, which assumes that people approximate optimal solutions through a limited set of mental simulations, we explain common heuristics as small sample approximations. Moreover, we assess how participants re-use computations and how they learn to approximate inferences over time.