Abstract
Sensory information integration is often near-optimal (reliability-weighted averaging, e.g. Ernst and Banks, 2002, Nature), but recent reports of suboptimality emphasize the need for process models accounting for suboptimal behavior (Rahnev and Denison, 2018, BBS). In a previous study, when combining sensory and prior information, suboptimal behavior was better fit by a model weighting information according to intrinsic noise only (internal to the observer, e.g. sensory noise) rather than one that accounted for both intrinsic and extrinsic noise (external to the observer, e.g. a stochastic cue) (Kiryakova et al., 2019, BioRxiv). We explicitly tested this hypothesis in a sensory cue-combination task. Twenty participants used auditory (400ms white noise burst) and visual (4-dot cloud) cues to find a hidden virtual object. Dots were drawn from one of two Gaussian distributions and centered on the true location (low/high variance intrinsic-only cues). Error distributions for visual cue-only trials were used for participant-by-participant calibration, where we added extrinsic noise to the low variance condition (shifting the dot cloud) such that intrinsic+extrinsic and intrinsic-only cues were equally reliable. Our hypothesis predicts near-optimal behavior in intrinsic-only trials but suboptimal behavior in intrinsic+extrinsic trials. We tested this by comparing the weight placed on visual cues (obtained by regressing visual bias against visual-auditory conflict) to the optimal prediction (based on single cue performance). Weight placed on the intrinsic-only cue did not differ significantly from optimal (p = .351). Weight placed on the intrinsic+extrinsic cue was significantly higher than optimal (p < .001), but significantly lower than an optimal prediction that ignored extrinsic noise (p = .014). Our results show that under-accounting for extrinsic noise can cause suboptimal perception and decision-making. Human perceptual systems may be well equipped to account for intrinsic uncertainty, but accounting for extrinsic uncertainty is more challenging, possibly in part because this uncertainty must be learned.