Abstract
Between ten and twenty years ago a broad range of perceptual learning (PL) results featured the lack of transfer across locations, tasks and features, implicating changes in early visual cortical areas. It held out promise for a simple unifying model of PL. In subsequent years the story has grown more complicated. Novel experimental protocols, many of which are summarized in Sagi's review “Perceptual Learning in Vision Research” (Vision Research (2010), doi:10.1017/j.visres.2010.10.019 or go to PubMed), show that learning can surprisingly transfer to untrained locations. Our recent results show that the learning can be transferred to untrained retinal locations and features even if an irrelevant task was used in one arm of the double training protocol. Competing principles have been formulated by several researchers for understanding this new data, but a concrete computational model to actually predict the data is lacking. The computational models that have been developed are generally used to predict the modelers own restricted data. In physics, where acquiring data is often costly, the researchers have come together to share in collecting critical data. This enables the computational modelers to test and distinguish among the competing models. The perceptual learning community has recently adopted this approach to collect a large trustworthy dataset for comparing models. At the 2010 Perceptual Learning Workshop in Israel, the Modelfest perceptual learning initiative was launched. The group considered not only what would be critical experiments, but also hidden methodological and statistical issues. This presentation reports on the current status of the ModelfestPL group's effort. In addition we will open up discussion for the next stage of the ModelfestPL deliberations.