Abstract
Over the past decade, many studies have used mixture models to interpret working memory data, drawing a distinction between capacity (number of items) and resolution (precision of representations) (Zhang & Luck, 2008), or proposing a mix of different precision memories (van den Berg et al. 2012). The results have led to numerous influential claims about the nature of working memory and long-term memory. Here we show that this entire class of models rely on erroneous assumptions about psychological similarity space and that once this is taken into account, no mixture model is required. Rather than 3 parameters (guess, precision, variability of precision), we equally accurately characterize the distribution of responses from continuous report using only a single free parameter (memory strength). The crucial insight is that while the color space used in these studies is perceptually uniform when comparing nearby colors, the distance between the target color and the foils range from small (1deg) to very large (180deg). This introduces well-known non-linearities in confusability for items with distances that vary considerably in physical space. In particular, we find an exponential-like fall-off in confusability with distance in 2D color space. After taking this non-linearity into account, a basic equal-variance signal detection model fits nearly all existing working memory studies. Our model also correctly makes novel predictions, including how the shape of the response distribution changes across different choices of color wheels and how the best fit parameters of mixture models differ when fitting orientation data rather than color data. This suggests a major revision of previous research, as it shows that the distinction between capacity and resolution is illusory. It also more directly relates previous models of long-term memory to the study of working memory; and offers a new theoretical framework for understanding the content and structure of working memory.
Meeting abstract presented at VSS 2018