Abstract
Although we have remarkable insight into the variations in quality of our visual working memory (WM) representations (Rademaker et al, 2012), how this uncertainty arises from neural activity patterns remains unknown. Bayesian theories of probabilistic population coding posit that information is represented as a probability distribution over feature values within populations of noisy neurons, and that the width of distributions directly indexes the uncertainty with which a feature is represented (Pouget et al, 2000; Ma et al, 2006; Jazayeri & Shadlen, 2006). Previous efforts to relate behavioral performance to neural WM representations (Ester et al, 2013; Sprague et al, 2014; 2016) have used linear methods, which cannot utilize the noise in neural responses to optimally constrain decoding. This is a critical challenge, as noise places critical constraints on representations of information in neural activity patterns (Averbeck et al, 2006). Here, we adapted a recently-published decoding method to measure representations of spatial positions in WM, as well as their uncertainty (van Bergen et al, 2016). This method, based on a Bayesian generative model of neural activity which incorporates spatial preferences of individual voxels and estimates of their noise, results in a full likelihood function over feature values, rather than a point estimate. Participants remembered a precise spatial position over an extended delay interval (10 s) while we imaged cortical activation patterns using BOLD fMRI. Decoded likelihood functions from visual cortex yielded accurate estimates of decoded feature values (mean of the likelihood function). Furthermore, the uncertainty of feature representations (circular standard deviation of the likelihood function) accurately reflected the noise of the representation: on trials with greater uncertainty, decoding error was higher. These results support variable precision models of WM, which posit that items are maintained with different levels of precision across items and across trials (van den Berg et al, 2012).
Meeting abstract presented at VSS 2017