[Paper Review] A flexible Bayesian method for adaptive measurement in psychophysics

Simon Barthelmé, Pascal Mamassian|ArXiv.org|Sep 2, 2008

Statistical and numerical algorithms11 references3 citations

TL;DR

This paper presents a flexible Bayesian adaptive method for measuring psychophysical functions, focusing on threshold, slope, or full function parameters using entropy-based stimulus selection. It leverages Bayesian updating and mutual information maximization to minimize trials while maintaining high precision, with a complete MATLAB implementation for practical use in psychophysics research.

ABSTRACT

In psychophysical experiments time and the limited goodwill of participants is usually a major constraint. This has been the main motivation behind the early development of adaptive methods for the measurements of psychometric thresholds. More recently methods have been developed to measure whole psychometric functions in an adaptive way. Here we describe a Bayesian method to measure adaptively any aspect of a psychophysical function, taking inspiration from Kontsevich and Tyler's optimal Bayesian measurement method. Our method is implemented in a complete and easy-to-use MATLAB package.

Motivation & Objective

• To develop a general-purpose Bayesian adaptive method for measuring any aspect of a psychometric function, including threshold, slope, or full parameter set.
• To minimize experimental time and participant burden by selecting stimuli that maximize information gain at each step.
• To provide a theoretically grounded, computationally efficient framework using Bayesian inference and mutual information.
• To offer a ready-to-use MATLAB package to lower the barrier for adoption by researchers without advanced programming or statistical expertise.
• To improve upon traditional methods like the method of constant stimuli and staircase procedures by dynamically adapting to observer responses and reducing wasteful testing.

Proposed method

• The method uses Bayesian updating to iteratively refine beliefs about psychometric function parameters using Bayes' theorem, starting from a prior distribution informed by prior knowledge.
• Stimulus levels are selected to maximize the expected information gain, measured as the reduction in entropy of the posterior distribution over parameters.
• Mutual information between stimulus-response pairs and parameter estimates is computed using kernel density estimation for non-parametric conditional density estimation.
• The approach is extended to handle transformations of parameters (e.g., thresholds at different performance levels) via invariance of mutual information under one-to-one transformations.
• For computational efficiency, the method incorporates Laplace approximation and fast algorithms such as those based on the Fast Fourier Transform and tree-structured kernel density estimation.
• A complete MATLAB package is provided, enabling researchers to implement the method without writing custom code.

Experimental results

Research questions

• RQ1 How can Bayesian adaptive methods be generalized to estimate not only thresholds but also slope and lapse rate of a psychometric function efficiently?
• RQ2 What stimulus selection rule maximizes information gain while minimizing the number of trials in psychophysical experiments?
• RQ3 How does the proposed method compare to traditional adaptive methods like QUEST or ZEST in terms of accuracy and efficiency?
• RQ4 Can mutual information between stimulus-response pairs and parameter estimates be reliably estimated in high-dimensional parameter spaces using non-parametric techniques?
• RQ5 To what extent does the method remain robust when estimating derived parameters (e.g., thresholds at 75%) rather than raw parameters (μ, σ, λ)?

Key findings

• The proposed Bayesian adaptive method significantly reduces the number of trials required to estimate psychometric function parameters by selecting stimuli that maximize expected information gain.
• Entropy reduction in the posterior distribution serves as a reliable proxy for information gain, enabling efficient and theoretically sound stimulus selection.
• The method maintains high accuracy even when estimating complex features such as threshold or slope, outperforming traditional methods like the method of constant stimuli that waste trials on uninformative stimuli.
• Kernel density estimation and mutual information computation allow non-parametric estimation of information gain, making the method applicable to a wide range of psychometric function shapes.
• The implementation of the method in a user-friendly MATLAB package enables broad adoption by researchers without advanced statistical or programming expertise.
• Theoretical and empirical results confirm that mutual information is invariant under one-to-one transformations of parameters, justifying the use of threshold-based information measures even when estimating from transformed parameters.