Conference on Cognitive Computational Neuroscience

Animals learn to make accurate decisions over long periods of time. In their journey from naïve to expert, animals often exhibit diverse trajectories, learning to make decisions in different ways. The behavioral, neural and computational principles underlying these learning trajectories are not yet understood. We longitudinally recorded dopamine (DA) release in dorsolateral striatum in tens of mice while they learned a psychophysical visual decision-making task from naïve to expert. Mice developed diverse learning trajectories over several days. Striatal DA responses to visual stimuli and rewards strongly reflected learning trajectories, showing substantial differences across mice. To gain computational insights, we designed a deep linear neural network model trained with gradient descent. The model reproduced the different learning trajectories, revealing that differences are largely determined by each animal's initial weighting of visual stimuli in the decision process, relative to their bias. The model also accounted for DA signals throughout learning; the dynamics of its prediction errors closely resemble DA signals across mice. Our work reveals computations that could underlie diverse learning trajectories, and demonstrates the dopaminergic representation of these computations.