Bayesian priors are encoded independently from likelihoods in human multisensory perception

Area of Science:

• Cognitive science
• Computational neuroscience
• Bayesian modeling

Background:

• Human crossmodal information combination aligns with optimal Bayesian causal inference.
• Bayesian priors are intuitively considered independent of current stimuli.
• The invariance of priors to varying likelihoods in Bayesian computation remains empirically unaddressed.

Purpose of the Study:

• To empirically investigate whether prior probabilities in Bayesian models are independent of likelihoods.
• To examine the stability of priors when faced with significantly altered stimulus information.

Main Methods:

• Manipulation of presumed likelihoods using two distinct sets of stimuli.
• Estimation of prior probabilities under different likelihood conditions.
• Empirical investigation of prior-stimulus independence.

Main Results:

• Estimated prior probabilities were found to be independent of the immediate sensory input (likelihoods).
• This independence suggests that priors remain stable despite changes in stimulus characteristics.
• Findings support the notion of stable internal representations in Bayesian inference.

Conclusions:

• The study provides empirical evidence for the independence of priors from likelihoods in human Bayesian computation.
• This supports a key characteristic of Bayesian models where priors are not influenced by immediate sensory data.
• Results clarify computational principles in crossmodal integration and segregation.