Abstract
Previous studies showed that both block feedback and trial-by-trial feedback can facilitate perceptual learning (Herzog & Fahle, 1997). However, it is not clear how they facilitate learning similarly or differently. Using an asymmetrical set of vernier stimuli (-15'', -10'', -5'', +10'', +15'') that contain more left than right offsets that has been shown to introduce response bias (Herzog & Fahle 1999; Herzog, Eward, Hermens & Fahle, 2006), Aberg and Herzog (2012) compared different forms of feedback and argued that both trial-by-trial feedback and block feedback support improvements in sensitivity, while only trial-by-trial feedback induces criterion shifts. Here we provide a comprehensive model for their complex results. Using the AHRM (augmented Hebbian Reweighting Model, Petrov, Dosher & Lu, 2005, 2006; Liu, Lu & Dosher 2010, 2012), we successfully modeled both overall performance change and behavioral shifts in response bias in the Aberg & Herzog (2012) data, including no feedback, trial-by-trial feedback with and without reversed feedback for -5'', and similar forms of block feedback. The Hebbian learning algorithm incorporates trial-by-trial feedback, when present, simply as another input to the decision unit and uses observer's internal response to update the weights otherwise. Block feedback alters the weights of the bias correction unit in the model (Liu, Dosher, & Lu, 2013). In the AHRM simulation, the response bias, or criterion shift, is induced by training on the asymmetrical set of vernier stimuli with trial-by-trial feedback, also influenced by reversed feedback. Training the asymmetric set incorporates biases into the weights between representation and decision. Block feedback counterbalances response bias with adaptive criterion control leading to less bias in block feedback conditions. The AHRM provides a detailed quantitative account for the results in Aberg & Herzog (2012), and makes generative predictions about differential training sets and feedback in perceptual learning.
Meeting abstract presented at VSS 2014