A dual-pathway dysregulation in prefrontal-habenular circuits mediates stress susceptibility.

One of neuroscience’s greatest achievements is demonstrating the strong correlation between midbrain dopamine activity and reward-prediction errors (RPEs), the learning signals that drive reinforcement algorithms in machine learning. RPE correlates, identified in a range of contexts and species, have led to the leading theory that dopamine encodes RPE and that basal ganglia circuits implement reinforcement learning. However, a growing amount of experimental evidence shows exceptions to RPE/dopamine correlations. Dopamine spikes in response to aversive and novel stimuli and expectations about information, not just reward. Aversive, novel, and informative events do not carry a positive RPE but are key for responding to changes in the environment. Important recent evidence shows that dopamine also tracks the rate at which rewards are expected. Dopamine is also known to perform roles outside of learning. Beyond just signaling, the circuit actively controls our actions from one moment to the next, including moment-to-moment movement and decision biases. In doing so, it has been shown to control the “explore-exploit tradeoff”, a key reinforcement learning concept where we decide whether to try new approaches or stick with what is known. Meanwhile, the striatum, a primary target of dopamine, performs a form of dimensionality reduction, filtering cortical inputs into a compact set of task-relevant information.