Planning and navigation, abilities fundamental to animal life, rely on elaborately structured circuits and systems in and around the hippocampal formation. In these regions, the responses of many spatially-tuned cell types—among them, grid, place, time, border, head direction, and speed cells—enable cognitive maps of the environment to form, and facilitate goal-directed navigation. How can we understand the logic and architecture of these circuits? To assist planning and navigation, they must predict trajectory-dependent future states based on noisy, occluded sensory input. To understand how they function, we train a recurrent neural network with biologically grounded constraints and segregated sensory and proprioceptive (internally-sensed motion) inputs to predict the next sensory signal. The network develops internal units displaying responses of key cell types of the hippocampal formation and reproduces much of their empirical phenomenology, including rapid "remapping" in previously learned environments and forward-planning sweeps during navigation. Such machine-learning based approaches offer an alternative to conventional mathematical modeling of physical phenomena.
Zoom link: https://icts-res-in.zoom.us/j/95178801699?pwd=YGl9bkdM4NJ3YNOV25OwCuWa5UNFj3.1
Meeting ID: 951 7880 1699
Passcode: 090909