A Two-dimensional Bump Attractor Model of Spatial Working Memory

Danielle Rager, Carnegie Mellon University

Primates are able to preserve knowledge of a visual cue, even during a delay period in which the cue is removed, using the persistent firing activity of prefrontal cortex (PFC) neurons that are spatially tuned to the cue. Model networks have replicated the stable "bump" of activity observed from PFC neurons with similar orientation preferences by creating a translationally symmetric, continuous "ring" of attractor states, corresponding to the possible angular locations at which a visual cue can be displayed in relation to the primate's central fixation point. Such model networks are able to encode a persistent representation of a visual stimulus in one-dimensional space. Using a novel spatial working memory task, in which primates must recall not only the orientation of a visual cue, but also its radial distance from the central fixation point, we demonstrate that PFC neurons are preferentially tuned to a two-dimensional location in polar space. Furthermore, we show that during such a task, populations of PFC neurons exhibit high-dimensional firing rate activity. Specifically, sub-populations of PFC neurons with tuning preferences for opposite visual hemifields have anti-correlated activity. We develop a new "bump attractor" model that is capable of encoding a persistent two-dimensional spatial representation of a visual cue. Our model consists of a spiking network of recurrently connected excitatory and inhibitory neurons, in which the probability of synaptic connection is a function of cell type and the two-dimensional distance between the preferred spatial locations of the cells. We show initial evidence that such a network is capable of producing the high-rank co-variability structure observed in our PFC data.

Abstract Author(s): Danielle M. Rager, Sanjeev B. Khanna, Matthew A. Smith, Brent Doiron