Adaptive, Goal-Oriented Navigation Using a Model of Directionally-Polarized Place Cells

CA1 place cells in the hippocampus have been shown to exhibit directional tuning properties, forming vector fields pointing towards locations in the environment known as ConSinks (Ormond & O’Keefe, 2022). We present a model, inspired by these findings, for learning goal-oriented navigation tasks. Our model employs a population of place cells that develop directional preferences, and are updated via a novel reward-modulated learning rule that refines directional turning of individual cells based on experience. Agents using this model navigated to goals significantly faster and more reliably than state-of-the-art Reinforcement Learning algorithms such as Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO). We also demonstrate adaptation to new goals in a manner consistent with experimental findings, where the mean ConSink location shifts towards the new goal after it is introduced. Further experiments show that the model performs well with both goal-directed and random initialization of directional sensitivity, and that place cell density enhances learning efficiency. These results suggest a functional role for directional place cells in complex and obstacle filled environments.