More than Just Measurement Models? Processing Insights from Population Coding Models of Visual Working Memory Retrieval

Recent studies of visual working memory (VWM) have investigated a population coding framework, in which items in memory are coded by an array of noisy, tuned detectors. Within this framework, heavy-tailed distributions of memory errors found in continuous-outcome VWM tasks reflect noise in the detector array rather than guessing. Studies have compared alternative tuning functions and detector noise distributions, finding differences in fit in some cases and substantial model mimicry in others. We used population coding models to look for evidence of a sample-size model of memory capacity, which predicts that memory strength will follow an inverse square-root law in set size, reflecting changes in the neural resources available to represent items. In two experiments, we independently manipulated the difficulty of the memory and decision tasks by varying the set size and the color saturation of the stimuli and the response wheel, respectively. We compared four population coding models and found that: (1) a von Mises tuning function provided better model fit and more stable estimates than a similarity-based, exponential-decay tuning function; (2) normal and Gumbel noise functions provided comparably good fits, but different patterns of parameter estimates; (3) Depending on task complexity, the amplitude of tuning function followed either an inverse square-root law or power law in set size, as predicted by the sample-size and attention-weighted sample-size model, respectively. Our results provided evidence for the sample-size account of VWM capacity in continuous outcome tasks and highlighted the potential of the models to be more than just measurement models.