A General Framework for Inferring Bayesian Ideal Observer Models from Psychophysical Data.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Psychophysics

Background:

• Understanding how the brain transforms sensory input into perception is a key neuroscience question.
• Bayesian ideal observer models link sensory data and theory, but inferring the statistical prior from behavior is challenging.
• Current methods often assume simple Gaussian priors, limiting flexibility in modeling complex sensory environments.

Approach:

• Reviews the general problem of inferring priors from psychophysical data.
• Introduces a novel approach using Gaussian mixture models to parameterize arbitrary prior shapes.
• Develops an analytical solution for psychophysical quantities, enabling numerical optimization to recover prior shapes.

Key Points:

• The statistical prior in Bayesian models cannot be directly measured and must be inferred.
• A flexible method is needed for priors not well-approximated by simple functions.
• Gaussian mixture models offer a flexible way to represent complex prior shapes.

Conclusions:

• The proposed method provides a flexible analytical framework for inferring arbitrary Bayesian prior shapes.
• This approach enhances the ability to model how prior knowledge influences sensory perception.
• A MATLAB toolbox is available to implement the described computations.