Abstract
Visual Working Memory (VWM) functions as an interface between higher cognition (e.g., reasoning, decision making) and perception. As such, we predict that the representations of VWM will have both a continuous component (i.e., to interface with continuous feature spaces in perception) and a discrete/symbolic component (i.e., to bind multiple features into a structured representation). To explore this hypothesis, we sought to manipulate both discrete binding and continuous features. We presented participants with a rapid serial visual presentation (RSVP) paradigm of two adjacent squares changing colors either monotonically (continuous condition - changing in one direction along the color wheel), or randomly (random condition - changing randomly across values of the color wheel). Participants were asked to report the colors of a briefly cued target-pair from the middle of the sequence. To minimize verbal coding of the colors, participants performed verbal-shadowing of a steam of words which they were required to repeat out loud. We predicted a larger bias (i.e., shift from the true target color) in reporting the target colors in the continuous condition relative to the random condition and a control condition - because of lag under continuous change. In contrast, we predicted more swaps (i.e., reporting its counterpart’s color for a color in the pair) in the random condition than in the continuous condition and the control condition - because random changes will disrupt binding. Consistent with our predictions, mixture model fitting revealed a larger bias in the continuous condition and a repeated-measure ANOVA showed more swaps in the random condition. This pattern of bias and swaps would not have emerged if VWM lacked either continuous or discrete representations respectively. Hence, we suggest that VWM representations consist of two different components — continuous representations for individual features and discrete symbolic representations for the bindings of features.