Abstract
Perceptual learning is a multifaceted process that may involve general learning, between-session forgetting or consolidation, and within-session rapid relearning and adaptation (Yang et al., 2022). The traditional learning curve, often derived from aggregated data in blocks or sessions comprising tens or hundreds of trials in most perceptual learning studies, may have obscured certain component processes. In a previous study, we developed three non-parametric inference procedures to estimate fine-grained contrast threshold learning curves in a Gabor orientation identification task, measured with the staircase procedure. In this work, we introduce a non-parametric Bayesian inference procedure to estimate the posterior distribution of the block d' learning curve in Yes-No tasks measured with the method of constant stimuli, incorporating varying block sizes. The model assumes the decision criterion as a constant likelihood ratio across all blocks for each subject. We applied the method with three block sizes (10, 35, and 100 trials/block) to a global motion same-different judgement task conducted over 3500 trials across five sessions (Yang et al., 2022). The goodness of fit to the data increased with the temporal resolution of the analysis. Model comparisons, based on the Bayesian Predictive Information Criterion (BPIC), identified the 10 trials/block model as the best fit. When fitting a multi-component generative model of perceptual learning (Zhao et al., submitted) to the average d' learning curves at the group level, we uncovered general learning, between-session forgetting and within-session rapid relearning with 10 and 35 trials/block. In contrast, the original study with 100 trials/block only identified general learning and within-session rapid relearning. The non-parametric Bayesian inference procedure offers a versatile framework for high-temporal resolution assessment of the component processes in perceptual learning across diverse tasks and testing paradigms.