Characterizing the impact of category uncertainty on human auditory categorization behavior

Adam Gifford and Alan A Stocker and Yale E Cohen
CoSyNe Computational and Systems Neuroscience, Salt Lake City UT, February 27 - March 02 2014
Poster presentation.

Categorization is an essential process in auditory perception. However, proper categorization is often challenging because categories often have overlapping boundaries. Here, we explore the degree to which human subjects are able to learn and use category distributions and their prior probabilities, and the computational strategy they employ to make use of this prior information when performing category decisions. In a set of experiments, we asked subjects to classify a tone burst into two categories according to the perceived tone frequency. Tone stimuli were sampled from two uniform category distributions that overlapped in tone frequency. In different blocks of trials, we systematically varied the prior probability of presenting a tone from one versus the other category. We found that subjects consistently learned these changes in prior probabilities and appropriately shifted their decision boundary toward the category with higher probability. To test the computational strategy that governed subjects' decision behavior, we first characterized subjects' individual uncertainty levels in the tone stimuli using a two-tone discrimination experiment. We then generated parameter-free predictions for individual subjects' categorization behavior for a Bayesian model with either an optimal decision strategy or a strategy that selects decisions according to the category posterior probability (i.e., probability matching). Indeed, we found that the model with the probability-matching strategy was a substantially better predictor of behavior than the model with the optimal strategy. Moreover, this finding held even when fitting the model to the data. Our work provides evidence that human observers apply a probability-matching strategy in categorization tasks with ambiguous (overlapping) category distributions. Furthermore, the model fits confirmed that human subjects were able to learn category prior probabilities and approximate forms of the category distributions.