Zoom Webinar URL: https://mit.zoom.us/j/97292217964
Humans make complex inferences from sparse observations, and rapidly integrate new knowledge to control their behavior. These abilities result from systematic organization of knowledge in the brain, called a cognitive map. The hippocampal-entorhinal (HC-EC) system is known to be important in the construction of the cognitive map in continuous domains, and may also contribute to the organization of discrete relational knowledge in humans. In this talk I will focus on the spatial domain, proposing that spatial knowledge may have a globally topological and locally metric organization, thus allowing fragmentation of space into submaps. I will present proposed human behavioural experiments for probing whether such submaps drive fast learning in complex spaces through compositional representations i.e., representations that encode space in terms of recombinable submaps. Next I will talk about principles that may guide fragmentation of space into submaps and present a plan to build an algorithmic model of map fragmentation. Finally, I will present a framework for building a neural circuit model of map fragmentation using a randomly mixed modular grid-place cell network that exhibits exponentially many robust fixed points. The network generalizes its stored inputs to create stable EC-HC attractor states around every pattern in the grid coding space, despite training over a vanishing fraction of the grid coding space.