Maintenance and Metalearning of Time Interval Representations

Abstract

When we perform actions in the world, we estimate what is happening around us. That information goes through a series of transformations in the brain in order to execute an action that meets our goals. For example, we might remember the speed of a car in order to decide when to cross the road. These transformations can be simple, for example based on physics models of speed and time, like the car example, or they can be complex and built around evolutionary and experience-based statistical regularities in the world. This thesis uses a sensorimotor time production task to investigate different types of transformation and noise that exist between observation and action. First, I will propose a task which utilizes memory of a time interval in order to probe memory noise, memory storage, and inference over internal noise. To do this, monkeys perform a delayed time reproduction task. I find that the behavior is consistent with the the brain storing the memory as a function of time, and that the inference does not mitigate the internal memory noise. Second, I investigate how estimated prior distributions change when the statistical regularities of the world change. Monkeys perform a blocked time reproduction task, and behavior across policy transitions shows fast adaptation to new policies. I apply this algorithm to a model and fit it to behavioral data. Third, I display some preliminary neural data gathered during these tasks as well as hypotheses for neural implementation. With these experiments, I utilize a simple task to pick apart transformations that occur between observation and action.