One-shot learning of paired association navigation with biologically plausible schemas

Schemas are knowledge structures that can enable rapid learning. Rodent one-shot learning in a multiple paired association navigation task has been postulated to be schema-dependent. But how schemas, conceptualized at Marr's computational level, correspond with neural implementations remains poorly understood, and a biologically plausible computational model of the rodent learning has not been demonstrated. Here, we compose such an agent from schemas with biologically plausible neural implementations. The agent contains an associative memory that can form one-shot associations between sensory cues and goal coordinates, implemented with a feedforward layer or a reservoir of recurrently connected neurons whose plastic output weights are governed by a novel 4-factor reward-modulated Exploratory Hebbian (EH) rule. Adding an actor-critic allows the agent to succeed even if an obstacle prevents direct heading. With the addition of working memory, the rodent behavior is replicated. Temporal-difference learning of a working memory gating mechanism enables one-shot learning despite distractors.