Purchase this article with an account.
Elyse Norton, Michael Landy; Dynamic estimation of prior probabilities in an orientation-discrimination task. Journal of Vision 2016;16(12):549. doi: 10.1167/16.12.549.
Download citation file:
© 2017 Association for Research in Vision and Ophthalmology.
Optimal sensory decision-making requires the combination of uncertain sensory signals with prior expectations. Signal detection theory describes the effect of prior probabilities as a shift in the decision criterion (i.e., response bias). Previous studies typically vary category probability between blocks and assume a fixed criterion for each block. Can observers track sudden changes in probability? We determine how observers update the decision criterion as category probabilities change. Stimuli were ellipses with orientations drawn either from category A or B. For each category, ellipse orientation was drawn from a Gaussian distribution. The category distributions were partially overlapping with different means and equal variance. Observers performed two tasks in separate sessions. (1) Covert-criterion task: An ellipse was shown and the observer indicated the category by keypress. (2) Overt-criterion task: The observer adjusted the orientation of a "criterion line" and then an ellipse was presented and categorized based on the criterion line's orientation. Feedback was provided in both tasks. Before each session, observers were trained on the category orientation distributions (2AFC, equal category probabilities). For both tasks, category probabilities were updated using a sample-and-hold procedure. The probability for category A (p) was randomly selected from a set of five probabilities (range: [.20, .80]). The probability for category B was 1-p. Probabilities were updated every 80-120 trials. In both tasks, observers tracked changes in category probabilities and used these estimates to modify the decision criterion. We compared observer performance to the optimal "omniscient" observer (i.e., an observer that sets the decision criterion optimally and knows the exact category probability on every trial) and found that observers undercompensated for changing probability. We will compare several additional models of task performance ranging from a simple reinforcement-learning model to models that explicitly estimate category probability and use that information to set the decision criterion.
Meeting abstract presented at VSS 2016
This PDF is available to Subscribers Only