Emergence of dynamically reconfigurable hippocampal responses by learning to perform probabilistic spatial reasoning

Abstract

Navigation in natural environments is computationally difficult: Location errors from motion estimation noise accumulate over time, while landmarks can be spatially extended and often look alike, thus providing ambiguous data. The brain contains a number of spatially tuned neurons coding for various navigational variables, but current models do not explain how these circuits could implement navigational computations that involve non-trivial spatial reasoning. We show, using a function-first approach, that neural circuits trained to efficiently solve spatial reasoning problems with performance on par with sequential probabilistic strategies reproduce some key properties of hippocampal coding, including heterogeneous tuning, conjunctive tuning, and low-dimensional dynamics. In addition, the models predict the emergence of tuning to key latent variables that are neither present in the input data nor trained as the end result of the task, and exhibit a spontaneous dynamical reconfiguration of tuning across time during a task as the computational demands evolve, reminiscent of some of the more complex dynamics observed in the hippocampus including a switch between location and displacement coding modes. These results provide a new functional framework for understanding the rich phenomenology and potential capabilities of navigation codes in the hippocampus and associated brain areas.