Feature Distribution Learning with Complex Contours in Visual Search

Abstract

One idea of how our visual systems process the information obtained from our environment is with ensemble perception by means of summary statistics. Here we present findings from a study on whether higher-order statistical properties are encoded with complex contours, our hypothesis being that they would be encoded. We replicated the feature distribution learning (FDL) method described by Chetverikov, Campana, and Kristjánsson (2016), which exploits the priming of pop-out effect, and relies heavily on role-reversal effects. We used a visual search task in which eleven observers searched for an oddly formed shape among 35 distractors. The stimuli were various shapes derived from the Validated Circular Shape Space (Li, Liang, Lee, and Barense, 2019). Observers indicated whether the target was situated in the lower half or the upper half of the array. Our results resemble previous findings in some aspects, but not for all observers with regards to encoding the shape of the Gaussian distribution. Response times depend on role reversals and the shape of the preceding distractor distribution, indicating that observers encode the mean and variance of the distribution. The decrease in response times for both the Gaussian and uniform distributions can be described by a two-step function, corresponding to the probability density function (PDF) of the uniform distribution. However, comparing prespecified models to the observed data revealed that approximately half of the observers encoded the shape of the Gaussian distribution, while the majority encoded the uniform distribution. Our results might indicate that observers need more learning trials in order to encode the Gaussian distribution for complex contours. Additionally, too much of the display may have been in the observers’ periphery, resulting in difficulties when trying to get an overall impression of the distribution. A solution to this would be rerunning the experiment, putting the individual shapes closer together, possibly using slightly smaller shapes, and increasing the number of learning trials. We conclude that encoding feature distributions of complex contours is possible, but the experimental design and method may require several adjustments.