Neural circuit mechanisms underlying contingency learning

Abstract

Studies on animal learning have shown that the efficacy of both associative learning and responding depends on contingency, the extent to which the likelihood of an outcome changes by a presentation of a stimulus. However, how neural representations in the mesolimbic dopamine system and ventral striatum are modulated by contingency remain unclear. Chapter I of my dissertation focuses on this by examining mice trained in a Pavlovian conditioning degradation task. I observed that anticipatory licking for odors previously associated with rewards significantly diminished when unpredicted rewards were introduced but remained stable with predicted additional rewards. Furthermore, dopamine (DA) axonal signals in the Olfactory Tubercle (OT) and Nucleus Accumbens (NAc), measured using photometry, mirrored these behavioral changes. While dopamine signals are commonly thought to resemble the temporal difference (TD) error in TD learning models, previous studies suggested these models couldn't explain the contingency degradation phenomenon. However, our findings indicated that DA responses in various experimental conditions align closely with a TD learning model incorporating state transitions reflective of task structure. Furthermore, we showed that recurrent neural networks (RNNs) trained in the task using a TD learning framework recapitulate dopamine responses and develop state representations consistent with the task states, merely from observations. Based on these results, we provided a theoretical framework linking TD errors to contingency and causal learning Many studies of reinforcement learning in rodents have involved the olfactory system. Previous studies have indicated that neural activity related to the odor valence have been observed in the OT, a relatively understudied structure situated in both the olfactory pathway and the ventral striatum. In Chapter II, I examined how valence representations of odor cues evolve over time in the two distinct neuronal populations in the OT, and how stable these representations are under different conditions in contingency learning using 2-photon microscopy. As the idea of representational stability and drift are vigorously debated, this study will add valuable biological data to inform conceptual ideas of neural representations in the brain. Together, these results provided new insights into the neural circuits involved in the representation of contingency in the mammalian brain.