Arousal-related mediation of perceptual belief updating across auditory domains
Belief updating refers to the integration of prior beliefs with incoming evidence and guides decision-making under uncertainty. In response to surprising events, this process is thought to be modulated by the locus-coeruleus-noradrenaline (LC-NA) arousal system, observable via pupil dilations (PDs). Pertinent literature has mostly focused on conscious, high-level decision-making and estimation processes, while assuming that the same principles apply to low-level sensory perceptual decision-making and generalize across tasks, domains and modalities. To address some of these assumptions, we devised a novel perceptual discrimination paradigm, investigating behavior and PDs across auditory domains. Participants were presented with auditory sequences of randomized length at a rapid pace, changing intermittently between two latent states: acceleration vs. deceleration (temporal group, N = 25) and clockwise vs. counterclockwise movement (spatial group, N = 22). Under high uncertainty, participants continuously inferred latent states to report the final state per sequence. To extract per-stimulus estimates of PDs during sequences, we fitted a deconvolution-based general linear model to the continuous pupil traces with a free amplitude parameter reflecting ongoing PDs. A Bayesian observer model was fitted to participants' responses and used to estimate information gain and surprisal for every stimulus. Participants' performance and PDs showed strong sensitivity to the occurrence of change-points. Both computational variables significantly predicted PDs in both domains, with information gain outperforming surprisal in a model comparison. Further model comparison revealed significant preference for models excluding possible domain-specific effects over models including them, pointing towards a constant effect over domains. We conclude that behavior and associated PDs observed in our purely perceptual auditory task align with Bayesian principles of belief updating. The observed lack of domain specificity supports the assumed generalizability of belief updating.