This PhD project investigates why music is so universally enjoyed by testing the intrinsic reward for learning hypothesis. It proposes that music engages ancient neural mechanisms linking prediction, surprise, learning, and reward, driving listeners to prefer sounds that slightly challenge their internal models. Project 1 will model how musical prediction and learning dynamics shape listening choices using behavioral data and AI-based prediction models. Project 2 will examine neural correlates of this process with intracranial EEG, focusing on interactions between auditory cortex, hippocampus, and amygdala within the VTA–hippocampal loop. Together, these studies aim to clarify how sound transforms into pleasure and learning, advancing both theoretical neuroscience and personalized music-based interventions.

Across every human culture and era, music has played a key role in society. However, its evolutionary purpose remains elusive. What neural mechanisms link the individual processing of patterned sounds to the rewarding and social experiences that music generates? The near-universality of a seemingly unnecessary auditory stimulus suggests that understanding how music influences the brain may reveal fundamental principles of neural function and how these mechanisms motivate cultural behaviors.

At the individual level, the link between sound and reward is often hypothesized to rely on prediction. These theories propose that the surprise arising from an unexpected note or chord generates musical tension, which is then translated into a rewarding experience once resolved. Yet it remains unclear why prediction should drive reward, since predictive processing is normally associated with efficiency—seemingly at odds with the pursuit of unnecessary surprise.

Interestingly, both prediction and reward have been separately linked to a third domain: learning and memory. For example, abstract pseudoword learning has been causally linked to dopaminergic systems: generating reward from a process with no extrinsic value. Similarly, structured surprise has been shown to enhance statistical learning, sequence-order memory, and the chunking of memory events. These findings suggest that learning may be a key mediator connecting prediction to reward.

I therefore hypothesize that the drive to experience music is motivated by an intrinsic reward for improving our internal models—that is, for enhancing our understanding of our own experiences. This mechanism likely extends beyond music, serving as an evolutionarily useful principle of human–environment interaction: predictions trigger surprise; surprise enables learning; learning leads to reward; and reward consolidates learning into memory, improving future predictions. Music provides a spark that activates these deeply ingrained evolutionary mechanisms, encouraging a better understanding of our environment. A key principle of this theory concerns how listeners select music, suggesting that individuals seek music whose complexity lies just beyond their internal models. However, music-listening behaviors and musical navigation have not been well studied in an openly accessible manner.

The proposed PhD thesis will investigate the intrinsic reward for learning hypothesis in two complementary projects focused on musical exploration.

In Project 1, we ask which features of musical prediction, reward, and learning guide listeners’ choices to continue listening to or discard a specific piece of music. Answering this question will provide deeper insight into how abstract sound sequences generate rewarding experiences and may lead to AI models capable of self-curating training datasets. To address this question, we will study the music-listening behaviors of a large cohort of participants by collecting an online behavioral dataset. Listeners will navigate a musical database through a series of possible actions: continue listening, restart the current piece, skip to the next piece, or switch to a different genre.

We will use recently developed AI models, fit to participants’ reported musical histories, to estimate how the dynamics of prediction, learning potential, and experienced reward drive action selection. Our ultimate aim is to develop a new computational model of music selection that integrates these factors. We expect this model to be guided by principles of exploitation versus exploration: an initial exploration phase allows listeners to find a piece or genre consistent with their internal models; subsequently, they will select music that challenges and refines those models, maximizing reward. Over time, preference for a given piece or genre may wane as predictive models adapt to its statistical features and learning potential diminishes.

In Project 2, we will investigate the neurobiological correlates of this process. Previous research examining the intrinsic reward for learning hypothesis in other domains has identified the VTA–hippocampal loop as a key mechanism for dopamine release that consolidates learning. We will test the hypothesis that prediction dynamics optimizing learning trigger dopamine release in the hippocampus, thereby enhancing sequence memory for specific levels of surprise.

We will collect intracranial EEG (iEEG) data from epileptic patients with clinically implanted electrodes as they perform a pared-down version of the behavioral experiment—making individual choices to re-listen to 30-second clips, continue to the next part of a piece, or begin a new one. Our analysis will focus on interactions between auditory cortical processing and activity in deep brain structures—particularly the hippocampus and amygdala—which are frequently targeted in clinical settings. Using both our established AI models and the new choice-selection framework, we will investigate how surprise encoding in the auditory cortex translates into learning and reward signals in the hippocampus and amygdala, and whether activity in these regions can predict participants’ decisions to continue or skip their current music. We highlight these areas for their accessibility via iEEG depth electrodes and their close functional and anatomical connections within the VTA–hippocampal loop.

The findings from this project will deepen our understanding of the interface between sensory and pleasurable experiences—a question fundamental to human experience—by emphasizing learning as a key motivator of behavior. Furthermore, they may allow us to predict musical preferences based on individual listening histories, enabling the development of more effective, personalized music-based clinical interventions. Future work will aim to identify how perceptual, predictive, reward, and learning disorders alter engagement with music, ultimately guiding the tailored deployment of musical interventions to improve clinical outcomes.

The project will be housed within the Human and Artificial Perception (HArP) team at the Institut de l’Audition, a center of the Institut Pasteur. The HArP team comprises passionate and supportive researchers at all career stages, dedicated to developing explainable computational models that advance our understanding of how natural sequences such as speech and music are processed. The selected PhD candidate will join this team as well as neighboring groups in human cognitive neuroscience, contributing to a vibrant community engaged in both fundamental and translational research in the speech and hearing domains.