Abstract
The Augmented Hebbian Reweighting Model (AHRM; Petrov et al., 2005) has successfully modeled various phenomena in perceptual learning. Fitting the AHRM to data presents a significant challenge because, as a sequential learning model, it must be simulated to generate performance predictions with sequential trial-by-trial updates, and estimation of the AHRM parameters is generally done using hierarchical grid-search methods. In this study, we introduce three modeling technologies to facilitate AHRM fitting: A Hierarchical Bayesian AHRM (HBAHRM) that incorporates population, subject, and test levels to estimate the joint posterior hyperparameter and parameter distribution while considering covariance within and between subjects; vectorization techniques with PyTensor to drastically speed up simulations involving multi-dimensional arrays; and pre-computed the likelihood function of the AHRM. We fit the data from Petrov et al. (2005), which investigated perceptual learning in an orientation identification task with 13 subjects in two external noise orientation contexts. We found that the HBAHRM provided significantly better fits to the data than the Bayesian Inference Procedure that inferred AHRM parameters for each subject independently. At the population level, the HBAHRM generated fits with an Rsq of 0.852 and RMSE of 0.031 (in d’ units). In a simulation study, we found that the HBAHRM exhibited excellent parameter recovery and fit the simulated data with an Rsq of 0.982 and RMSE of 0.010 (d’ units). Additionally, the HBAHRM made excellent predictions of the performance of a new simulated observer with no data, 300 trials (all in one context), and 2700 trials (300 in one and 2400 in the other context) of data. The HBAHRM and the new modeling techniques can be readily applied to analyze data from various perceptual learning experiments and provide predictions of performance of new observers with no or limited data.