Bayesian adaptive stimulus selection for dissociating models of psychophysical data.

Area of Science:

• Computational neuroscience
• Perception and action modeling
• Psychophysics

Background:

• Comparing computational models is crucial for understanding perception and action.
• Current stimulus selection methods (predetermined or random) are limited in dissociating models effectively.
• Efficient and reliable model comparison requires stimuli that elicit differential model predictions.

Purpose of the Study:

• To expand the psi-algorithm for distinguishing between computational models, not just estimating parameters.
• To develop a more efficient and reliable method for stimulus selection in model comparison experiments.
• To test the algorithm's efficacy in dissociating sensory noise models and comparing target selection models.

Main Methods:

• Expanded the psi-algorithm to select stimuli adaptively for model comparison.
• Applied the algorithm to simulated ideal observers with different sensory noise models in a two-alternative forced-choice task.
• Validated the algorithm with human subjects in a speed perception task and in comparing target selection models under body acceleration.

Main Results:

• The enhanced psi-algorithm significantly improved the accuracy of model comparison compared to random stimulus selection.
• In human experiments, the algorithm rapidly converged to the correct sensory noise model, unlike random sampling.
• Adaptive stimulus selection demonstrated potential for stronger conclusions in comparing subtle effects, such as target selection under acceleration.

Conclusions:

• The enhanced psi-algorithm offers a more efficient and reliable approach to stimulus selection for model comparison.
• This technique overcomes limitations of predetermined and random stimulus selection, leading to more accurate model dissociation.
• The method is broadly applicable to various problems in computational neuroscience, perception, and action research.