Purchase this article with an account.
Sebastian Schneegans, Paul M Bays; Neural resource model explains visual working memory performance in whole-report tasks. Journal of Vision 2019;19(10):80b. https://doi.org/10.1167/19.10.80b.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Neural population models have successfully accounted for human visual working memory performance in a variety of cued recall tasks. These models assume that recall errors are caused by decoding from noisy population activity, and explain limits in memory capacity through normalization of total spiking activity across memory items. We apply this type of model to a new experimental paradigm, the analog whole report task, in which subjects must sequentially reproduce the remembered features of all items from a sample array, in an order either freely chosen or determined randomly. To explain recall performance in this setting, we incorporate two mechanisms into the model that have previously been used in different contexts. First, we propose that the total number of spikes contributing to the memory of each item is used as a measure of confidence in the decoded feature value, and that this determines the order of reports if it is freely chosen. Second, we assume that memorized feature values undergo random drift over extended delays, causing additional impairments in memory quality for the later reported items. We obtained close quantitative fits to response distributions for two different experimental conditions across five different set sizes and up to six different ordinal response positions, using only three free parameters. In particular, the model accounts for the continuous decrease of recall precision across ordinal response positions observed when report order is chosen by the subject. This aspect of the data has previously been interpreted as evidence for a fixed item limit in working memory. In this view, subjects can memorize only a subset of items at higher set sizes, and report those items first while producing pure guesses in their later responses. Our results demonstrate that a model based on continuous memory resources can explain the data without incorporating a fixed item limit.
This PDF is available to Subscribers Only