Cognitive Science Lunchtime Talk
Characterizing people’s priors over navigation task structure
Humans perform a large and diverse array of tasks (navigating, writing abstracts, etc.) with relative ease. Yet, if we consider all possible tasks one could formulate, we would not be good at most of them (e.g. playing Go). This simple observation leads to a basic, yet largely unanswered, question: what are the essential properties of tasks that our brain is good at solving? The curse of dimensionality suggests that task representations should be compact, filtering out redundancies. These reduced representations will manifest as a prior that reflects the structure of the tasks we have evolved to solve. In this project, we study these priors. We start with navigation tasks as these are easily mapped to graphs, which enables the quantification of task structure using graph theoretical tools. We use iterated learning (or Markov Chain Monte Carlo with People) - a process whereby an agent learns from data generated by another agent, who themselves learned it in the same way - to estimate participants' priors over task graphs. Our approach enables us to quantify what the human brain assumes about the abstract structures of graphs, and the individual variability in these priors. I will present simulations and preliminary results for priors over navigation graphs of university campuses.