Distilling noise characteristics and prior expectations in multisensory causal inference

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Modeling

Background:

• Human perception integrates multisensory information to reduce uncertainty.
• Bayesian observer models explain multisensory causal inference but often use simplifying assumptions.
• Traditional models assume constant sensory noise (homoskedasticity) and Gaussian priors.

Purpose of the Study:

• To challenge and relax assumptions of homoskedastic sensory noise and Gaussian priors in Bayesian multisensory perception models.
• To develop a flexible semiparametric approach for inferring sensory noise and prior shapes directly from data.
• To investigate the impact of these refined assumptions on understanding human multisensory perception.

Main Methods:

• Collected auditory-visual perceptual data from unisensory and bisensory tasks.
• Employed a semiparametric modeling approach to infer noise and prior distributions from participant data.
• Analyzed stimulus location estimates and same-different source judgments.

Main Results:

• Human sensory noise is eccentricity-dependent, plateauing in the visual periphery.
• Prior distributions feature a narrow central peak with smoother tails, deviating from Gaussian assumptions.
• Evidence of auditory range recalibration and increased sensory noise in multisensory conditions was observed.

Conclusions:

• Data-driven modeling reveals more complex sensory noise and prior structures than traditional assumptions.
• Findings necessitate revisions to Bayesian models of multisensory perception.
• The study provides new tools for perceptual causal inference research.