Fig. 1: Feedback in three-compartment HPC neurons enables context-sensitive representations.

a A 2D grid environment, in which the agent must visit a cue state (state 5) to receive a reward when it reaches the end state (state 9). Bottom, state transition 8–9 can be either rewarding (if preceded by 5) or punishing (if preceded by 7), leading to ambiguity in the value of the 8–9 transition. b Some models include a history-dependence a priori as part of their state representation. Here, the population activity vector is equivalent to the state vector at that time, so the two potential “8” states (shown by the dotted ovals), are disambiguated. c Alternatively, inference models compare internal predictions to external sensory observations, and update their internal models based on errors between their predictions and observations. Successfully trained inference models learn the latent structure of the task as part of their internal model. d Schematic of the hippocampal network model we propose. The network receives external inputs x\((t)\) into a basal dendritic compartment b\((t)\). Somatic activation s\((t)\) is a combination of basal activity b\((t)\) and recurrent apical feedback a\((t)\). e Schematic of the full model, including action neurons, which receive input from the representations in the hippocampal network and dictate the agent’s decisions in its sensory environment, providing a closed loop between the environment and the network.