Conference on Cognitive Computational Neuroscience

Learning a probability distribution from finite examples is a prerequisite for many perceptual and cognitive tasks. As an inverse problem, its difficulty lies in the potentially infinite size of the solution space. How does the brain handle the problem? Normative theories propose that the brain may build an internal model of observations using a mixture of simple clusters, the number of which could grow towards infinity. However, it is unclear how such an infinity may be reconciled with limited cognitive resources. Here, we first used novel density estimation tasks to obtain an explicit and detailed description of human participants’ internal models; then, we developed computational modeling methods to understand the cognitive processes underlying this high-dimensional behavioral set. We found that humans use overly complicated models for simple distributions but overly simple models for complicated distributions. The former arises from a preference for narrow clusters a priori to cover a wider range of observations, while the latter reflects a reduced willingness to continue introducing new clusters, leading to an economical model expansion that balances model complexity with cognitive cost.