Sequential Effect on Visual Classification: The Citrus Classification Paradigm

Abstract

Introduction: Prior knowledge affects performances in diverse perceptual and cognitive tasks (Caillies, S. et al., 2002; Johnson, E. J., & Russo, J. E., 1984), and human can adaptively learn mean and variance of the prior distribution (Berniker M., et al., 2010). In existing studies, however, samples from prior distribution were presented in random order, thus the effect of sequential order on the construction of prior distribution has not been explicitly tested. Here, we examined how the order of stimuli presentation affects the learning of prior distribution. Methods: Experiment consists of training and test sessions. In training session, which was to expose subjects to the prior distribution of stimuli size in a particular sequence implicitly, subjects were asked to classify fresh and rotten citrus according to the color. There were three conditions. In increasing or decreasing size conditions, the size of stimuli increased or decreased gradually with additional random fluctuation. In random size condition, the size of stimuli randomly varied. Each subject ran 630 trials in one of three conditions. Distribution of stimuli diameters was uniform between 1.4° and 4.3° in visual angle for all conditions. In test session, subjects classified 420 citruses of various sizes, which follow identical distribution used in training, into three size categories (small/medium/large) based on their own criteria. In both sessions, stimuli were presented at random position in the screen to prevent making virtual criteria for classification. Results: Subjects' internal criteria for classification in test session were modulated by the order of stimuli presentation in training sessions. The proportion of 'medium' responses was the lowest in increasing size condition and the highest in decreasing size condition, presumably reflecting the difference in estimated variance of stimuli distribution. Conclusion: We showed that the order of stimuli affect the learning of prior distribution and estimating criteria for classification using a novel experimental paradigm. Meeting abstract presented at VSS 2016