An attractor model explains space-specific distractor biases in visual working memory

Working memory (WM) enables retaining and manipulating information for brief periods. Attractor models have been developed for explaining diverse phenomena linked to WM, including error-correcting dynamics [1], cardinal repulsion [2], serial biases [3], and the like [4-5]. Yet, in the real-world, a relevant stimulus must be remembered despite interference from distractors that can occur unpredictably in time and anywhere in space. We provide novel experimental evidence, and suggest a computational mechanism, for biases induced by spatially and temporally unpredictable, task-irrelevant distractors in visual WM.
We tested n=24 human participants on a continuous response, visual WM task while recording brain activity with scalp electroencephalography. Briefly, participants memorized the orientation of bilaterally presented gratings following which they were retrospectively cued to the location more likely to be probed for response (70% cue validity). During the WM delay a distractor grating appeared at an unpredictable time and, randomly, with equal probability, at one of the two erstwhile stimulus locations. We discovered that the decoded neural representation of the cued stimulus was systematically biased towards the distractor’s orientation when the latter appeared in the same hemifield as the former (Fig. 1C, p=0.01). Yet, surprisingly, the cued representation was biased away from the distractor’s orientation when they appeared on opposite hemifields (Fig. 1C, p=0.026). A similar pattern of distractor-induced biases occurred for the uncued stimulus also (Fig. 1C; same-hemifield p=0.005, opposite-hemifield p=0.162). Behavioral reports confirmed these antagonistic bias trends (p<0.05, Fig. 1D).
We developed a WM model with distinct ring attractors [5] representing stimuli in each hemifield, while also incorporating topographic, cross-hemifield inhibition (Fig. 2A). The model faithfully replicated each of the experimentally observed biases, neural and behavioral (Fig. 2B-C).
In sum, task-irrelevant distractors induce space-specific, antagonistic biases in visual WM. Our results suggest that cross-hemifield inhibition suffices to fully explain these biases.