Bayesian adaptive stimuli selection for dissociating psychophysical models

Abstract

Comparing models allows us to test different hypotheses regarding the computational basis of perception and action. One difficulty in model comparison is that it requires testing stimuli for which the models make different predictions. To date, experiments contain typically a predetermined set of stimuli or sample randomly over a large range of stimulus values. Both methods have limitations; a predetermined set may not contain the stimuli that dissociate the different models and random sampling may be inefficient. To overcome these limitations, we derived an algorithm to efficiently dissociate psychophysical models using adaptive stimuli selection. Formally, our method selects stimuli that minimize the expected entropy of the posterior distribution across models after the next stimulus or stimulus pair has been presented. To test our algorithm, we considered the problem of comparing sensory noise models. Many computational models assume a specific perceptual noise type, like constant noise, Weber noise or a combination. The appropriate noise model can be inferred using a 2-AFC task. We simulated ideal observers with different noise models performing such a task. Stimuli were selected randomly or using our adaptive algorithm. On average the number of trials required to converge to the correct model was lower for the adaptive algorithm compared to random sampling. We also verified our algorithm in human subjects by inferring which of the aforementioned models underlie speed perception. Subjects were presented with two Gabor patches moving at different speeds and indicated which was faster. On a trial the speeds were either chosen randomly or by the adaptive algorithm. The adaptive procedure converged to the model reported in earlier work (Stocker & Simoncelli, 2006), whereas the random sampling method was often inconclusive. We conclude that our technique is more efficient and more reliable than the methods that are used to date to dissociate psychophysical models. Meeting abstract presented at VSS 2017.