Reward prediction refers to an organism's (or agent’s) anticipation of a reward following a cue, an action, or an event. It is a core concept in both behavioral psychology and computational models of learning. As an example, in cue-guided instrumental learning, animals learn over time that performing a specific action after a given cue is played will lead to a reward. Reward prediction is essential to form internal models of the world, including knowing what outcomes to expect given the current state (value function in reinforcement learning) or action (action-value function). This reward prediction signal is necessary to compute reward prediction errors (RPE), the difference between expected and actual reward, which is then used to update internal predictions. As such, reward prediction encodes what an agent knows, and can be seen as an instructive signal that guides future decisions and behaviors. Alternatively, RPE is considered a core learning signal used to update value functions (or reward predictions). In other words, prediction is used to act, and error is used to learn. Numerous evidences point to RPE being instantiated in the brain by the mid-brain dopaminergic circuit (ventral tegmental area and substantia nigra). Reward predictions is thought to be encoded in frontal cortical and striatal circuits, that includes the orbitofrontal, medial prefrontal and anterior cingulate cortices (OFC, mPFC and ACC), and the ventral and dorsal striatum (VS and DS). The amygdala, interacting with OFC for reward learning, also encodes associative value of cues.

In cue-guided, instrumental reward learning, sensory cues play a crucial role signaling the opportunity of accessing a reward. Contrary to the core reward prediction and value computation regions listed above, the sensory cortex is thought to reflect reward prediction by modulating sensory responses based on value information from elsewhere (e.g., amygdala, OFC). The thought is that the sensory cortex does not compute nor store reward predictions, but instead use those signals to enhance behaviorally-relevant stimuli to be prioritized by downstream regions. Contradicting this view, our previous work shows that more than enhancing perception, the auditory cortex (AC) directly encodes reward prediction, and that this signal is necessary for associative, instrumental learning (Drieu et al., Nature 2025). Reward prediction in the AC emerges extraordinarily fast, over only dozens of trials while task performance is still poor, and is initially delayed from cue-evoked activity. These findings challenge current conceptual models of learning-related plasticity in the sensory cortex. Further, we showed that the sensory cortex is essential for the acquisition of task contingencies and performance improvements in the task, but becomes dispensable at expert level, potentially tutoring subcortical structures that take over once the association are learned and can be correctly executed. This suggests conceptual shift in our understanding of the role of sensory cortex in cue-drive, instrumental reward learning, where the AC does not only prioritize relevant sound, but can instruct downstream regions for learning. 

This project aims at testing the properties of the AC reward prediction signal and elucidating the inputs necessary for its emergence. Notably, we address whether AC reward prediction encodes reward identity and negative outcome, and whether the OFC and the amygdala are responsible for its development.

Approach: head-fixed behavior in mice, two-photon calcium imaging, closed-loop optogenetics, large-scale data analysis. Learn more

This project assesses the contribution of spontaneous reactivation of experience-related activity in a distributed brain network in forming and consolidating reward-based memories. Neuronal reactivations have been primarily studied in the hippocampus in the context of spatial memory. Electrophysiological recordings in freely moving rats revealed that hippocampal neurons code for the location of the animal in the environment (‘place cells’). Surprisingly, hippocampal neurons activated during exploration are spontaneously reactivated during ‘offline’ periods, virtually reproducing the trajectory of the rat during its previous exploration. This replay occurs during transient hippocampal activity patterns called sharp wave-ripples (SWRs). Repetitive reinstatement of experience-related activity during high synchrony epochs provides ideal conditions for plasticity within the hippocampus and for communication with other brain regions. Moreover, SWRs and their associated replay are causally linked to memory functions. Interestingly, reactivations are found beyond the hippocampus, in prefrontal, sensory and associative cortical areas, as well as in subcortical structures such as dorsal and ventral striatum, ventral tegmental area, and amygdala. These extra-hippocampal reactivations and increases of hippocampal SWRs are reported in both hippocampus-dependent and -independent tasks, and extra-hippocampal reactivations and hippocampal SWRs exhibit a tight temporal relationship. This raises the compelling possibility that reactivations play a role beyond spatial memory, in stabilizing, forming, and consolidating memory traces from different memory systems. However, the role of reactivations in cue-guided, reward-directed learning remains largely unexplored. Filling this gap is of critical importance as it may identify how brain networks learn to link cues to rewarded actions, deepening our basic understanding, while also identifying a new mechanism for therapeutic intervention in substance use and memory disorders. It may also provide a united framework for how memory traces are formed and consolidated across different memory types. Due to technical difficulties in recording neuronal population activity across multiple brain circuits simultaneously, reactivations have been investigated in either one or two regions at the same time, limiting our understanding of the complex multi-structure interactions during reward-based behaviors. The goal of this project is to test the hypothesis that cue-driven reward learning relies on reactivation of task-relevant neurons distributed among multiple brain regions during the waking state, forming associative networks subsequently reactivated during sleep for consolidation. 

Approach: freely-moving behavior in rats, multi-site extracellular electrophysiological recordings, closed-loop neuronal manipulations, large-scale data analysis. Learn more