THE COMPUTATIONAL BASIS OF AESTHETIC REWARD: AUDITORY SEQUENCE PERCEPTION IN MICE AND IMPLICATIONS FOR MODEL-BASED REINFORCEMENT LEARNING

Computational theories of aesthetics posit that aesthetic value is a function of both immediate sensory fluency and long-term gains in predictive efficiency. This could account for phenomena like the preference for intermediate predictive complexity, repeated exposure, and the appeal of symmetry. We aimed to investigate the biological generality of this model by exploring how structural and semantic complexity drive intrinsic reward in the mouse.
We developed an automated 8-arm maze acting as a state machine. Spatial navigation into an arm triggered one of several sounds in real time. Each sound was linked to a specific arm in a block-by-block design, implementing a dynamical conditioned place paradigm. We generated synthetic auditory stimuli with varying sensory complexity, quantifying interval consonance/dissonance and amplitude modulation to vary acoustic roughness, and including semantically meaningful stimuli such as ultrasonic vocalisations (USVs). In additional experiments, we combined this behavioural assay with fibre photometry recordings in the Nucleus Accumbens core to monitor dopaminergic signalling as a proxy for value processing.
Consistent with the literature, mice exhibited a general preference for silence over synthetic stimuli, regardless of their consonance or structural complexity, but not over USVs. Unlike synthetic sounds, USVs drove a significant preference over the silent arm. Across individuals, exploratory drive correlated with the overall motivation to engage with unrewarded sounds.
These findings underscore the role of semantics as a driver of intrinsic reward. We are currently further investigating this role by associating abstract grammars with specific environmental contexts.