Latent Representations in Hippocampal Network Model Co-Evolve with Behavioral Exploration of Task Structure

Abstract

Real-life behavioral tasks are often complex and depend on abstract combinations of sensory stimuli and internal logic. To successfully learn these tasks, animals must pair actions or decisions to the tasks complex structure. The hippocampus has been shown to contain fields which represent complex environmental and task variables, including place, lap, evidence accumulation, etc. Altogether, these representations have been hypothesized to form a "cognitive map" which encodes the complex real-world structure underlying behavior. However, it is still unclear how biophysical plasticity mechanisms at the single cell level can lead to the population-wide evolution of task-relevant maps. In this work we present a biophysically plausible model comprised of a recurrent hippocampal network and an action network, in which the latent representational structure co-evolves with behavior in a task-dependent manner. We demonstrate that the network develops latent structures that are needed for solving the task and does not integrate latent structures which do not support task learning. We show that, in agreement with experimental data, cue-dependent "splitters" can only be induced at the single cell level if the task requires a split representation to solve. Finally, our model makes specific predictions on how biases in behavior result from experimentally testable biases in the underlying latent representation.