Abstract
In many circumstances, perceptual learning is tracked at a fairly coarse grain of measurement, often in blocks involving a hundred trials or more. Embedded in this paradigm is an implicit assumption that learning can be fully measured in the performance improvement across blocks. However, in most models of perceptual learning, learning potentially occurs after each trial. Such modeling calls for a re-examination of the trial-by- trial learning dynamics in perceptual learning experiments. In this study, we examine the data from a study that compared different forms of intermixed task training - "roving" of stimuli. Observers make orientation judgments (clockwise or counterclockwise) about sets of base angles drawn from {+/-12deg about -67.5deg, -22.5deg, +22.5deg, and +67.5deg relative to vertical}. Observers judge CW or CCW to 4, 2-similar, 2-dissimilar, or 1 base angles each trained in one of four locations, with an adaptive staircase tracking 75% correct. While the coarse block-wise analysis shows differences among groups in the first session, the trial-by- trial analysis shows the statistical equivalence of the groups at the beginning of training, followed by separation of the more difficult all-4 and 2-similar groups from the 2-dissimilar and 1 base angle group relatively early in the first session, followed by relative improvement of the 2-similar group. Combined with simulations of the specific staircase procedure used, the trial-by-trial analysis also reveals interesting micro-patterns such as within session learning and increasing lapse rate toward the end of a session. A simulation of the integrated reweighting theory (IRT, Dosher et al., 2013) using the same staircase procedure shows very similar micro-structure of contrast thresholds, as well as initial within-session settling of percent correct within the 90% confidence intervals of 75%. This analysis provides important information about the early phase of perceptual learning that is critical to distinguish different learning processes.
Meeting abstract presented at VSS 2018