Abstract
Visual working memory (WM) allows us to maintain arbitrary combinations of information bound into novel objects. The information comes from continuous feature spaces, resulting in limited precision and graded biases. WM theories are divided into two camps: slot-like models that include feature binding, and continuous attractor models that predict graded biases. No current models explicitly account for both these aspects of human WM. A recent computational model of WM models binding in a biologically plausible network. Rapid synaptic plasticity supports the formation of discrete-feature bindings and their maintenance in WM background, whereas persistent activity forms the WM foreground by holding a single representation in a focused state. In this study, we extend this model to represent continuous feature spaces. Bindings of multiple continuous features are stored as “plastic attractors” in flexibly-coding conjunction neurons and the network produces continuous report via pattern completion. Analysis of model activity shows that recall errors stem from competition between representations. Depending on the particular way the competition resolves in individual trials, simulated errors can be assigned to different error sources: imprecision of target representation, swap errors and random responses. The model accounts for a range of WM effects including the effect of set size on precision and misbinding, and serial order effects. It also produces biases generated by interference when multiple items are maintained concurrently and generates a novel prediction: similarity along the to-be-recalled dimension and along the probe-feature dimension should have different effects on the pattern of errors. Furthermore, continuous, plastic changes in conjunction neuron selectivity can account for serial biases, where recall is biased towards items encountered on previous trials. In summary, the continuous binding model can capture a variety of WM effects that other theories cannot, and shows that transient flexible coding can support binding of continuous features.