Bayesian active probabilistic classification for psychometric field estimation.

Area of Science:

• Psychology
• Machine Learning
• Signal Processing

Background:

• Psychometric functions are typically estimated using parametric models, with extensive research on unidimensional curves.
• Current methods for multidimensional psychometric fields are limited, especially for complex or easily parameterizable functions.

Purpose of the Study:

• To introduce a flexible, nonparametric Bayesian approach for estimating general psychometric fields.
• To develop an estimator that sequentially selects subject queries to improve estimation accuracy and efficiency.

Main Methods:

• Implemented a nonparametric Bayesian psychometric field estimator using Gaussian processes trained by active learning.
• Compared two actively sampled estimators against two non-actively sampled estimators using simulations of pure-tone audiograms.

Main Results:

• Actively sampled methods achieved accuracy comparable to non-actively sampled methods but required fewer observations.
• This efficiency advantage was consistent across various audiogram phenotypes representing human auditory perception.

Conclusions:

• Gaussian process classification offers a generalizable estimation procedure for multidimensional psychometric fields with binary responses.
• This approach shows significant potential for extending practical estimation to complex perceptual models previously inaccessible.