STRUCTURAL LEARNING OF REWARD PATTERNS IN THE PREFRONTAL CORTEX OF MICE

Learning and memory studies often focus on associations between pairs of stimuli, or stimuli and rewards. However, it is also possible to learn more abstract structures governing the reward environment. Here we test how mice rely on both strategies for learning, using a task that can tease them apart at the level of behaviour, and neural representation. Mice ran on a virtual corridor made of ten landmarks, and were rewarded at four locations #2,4,6,8 (‘goals’) (Fig. A). To learn the reward structure, one strategy is to learn the ten landmark-reward associations. An alternative is learning a structural pattern that predicts reward every second stimulus. Behaviour at the 10^th, unrewarded, landmark (‘test’) distinguished these strategies, as only the latter predicts responses there. Even after responses to all other non-rewarded landmarks stopped, mice continued responding to the 10^th landmark, consistent with structure learning, before learning the correct reward pattern (Fig. B). We recorded individual mPFC cells in task and examined neural activity as a function of progress towards different behavioural steps (Fig. C). Neurons initially tracked progress to the four goals, and the test landmark. Surprisingly, a subset of neurons maintained this alternating landmark representation even after learning the true reward structure (Fig. D). These findings suggest that even in simple stimulus-reward association paradigms, mice may favour a structural explanation for rewards, if one is available. Characterising representations of similar structural regularities is critical to understand how they could be used as building blocks for more complex structures.