Center-Surround Inhibition in Working Memory

Abstract

Summary Directing visual attention toward a particular feature or location in the environment suppresses processing of nearby stimuli [1]–[4]. Echoing the center-surround organization of retinal ganglion cell receptive fields [5], and biasing of competitive local neuronal dynamics in favor of task-relevant stimuli [6], this “inhibitory surround” attention mechanism accentuates the demarcation between task-relevant and irrelevant items. Here, we show that internally maintaining a color stimulus in working memory (WM), rather than visually attending the stimulus in the external environment, produces an analogous pattern of inhibition for stimuli that are nearby in color space. Replicating a well-known effect of attentional capture by stimuli that match WM content [7], visual attention was biased toward (task-irrelevant) stimuli that exactly matched a WM item. This bias was curtailed, however, for stimuli that were very similar to the WM content (i.e., within the inhibitory zone surrounding the focus of WM), and recovered for less similar stimuli (i.e., beyond the bounds of the inhibitory surround). Moreover, the expression of this inhibition effect was positively associated with WM performance across observers. In a second experiment, inhibition also occurred between two items simultaneously held in WM. This suggests that maintenance in WM is characterized by an excitatory peak centered on the focus of (internal) attention, surrounded by an inhibitory zone to limit interference by irrelevant and confusable representations. Here, thus, we show for the first time that the same center-surround selection mechanism that focuses visual attention on sensory stimuli also selectively maintains internally activated representations in WM.

Chunking as a rational strategy for lossy data compression in visual working memory

“…Ben-Yishai, Bar-Or, & Sompolinsky, 1995;Kohonen, 1982; Somers, Nelson, & Sur, 9 1995), and predict biases previously observed in visual working memory reports10 Almeida et al, 2015;Kiyonaga & Egner, 2016). We thus altered the Wei et almodel 11 to include broadly tuned inhibition in accordance with center-surround recurrence, 12whereby Feedback inhibition is stronger for neurons with similar color tuning 13 (figure 5B).…”

“…Thus in a sense, such bump collisions are analogous to the chunking implemented in our more abstract binary encoding model, yet they lack the selectivity necessary to mediate performance optimization, and indeed, predict the opposite pattern of performance than seen empirically, with worse performance for randomly spaced arrays and improved performance for fixed arrays (Wei et al, 2012) We considered whether other patterns of connectivity would remedy this issue. Notably, physiological data suggest that neural responses within such networks obey center-surround receptive field architectures that are present throughout the visual system (Hubel & Wiesel, 1959;1965), are supported by lateral connectivity (Ben-Yishai, Bar-Or, & Sompolinsky, 1995;Kohonen, 1982;Somers, Nelson, & Sur, 1995), and predict biases previously observed in visual working memory reports (Almeida et al, 2015;Kiyonaga & Egner, 2016). We thus altered the Wei et al model to include broadly tuned inhibition in accordance with center-surround recurrence, whereby Feedback inhibition is stronger for neurons with similar color tuning ( figure 5B).…”

“…If the two stored colors are similar enough to promote mutual recurrent excitation, each represented color will experience biased excitation in the direction of its neighboring color, and eventually the two color "bumps" will merge to form a single representation (figure 6A) (Wei et al, 2012). In contrast, if the stored colors are separated beyond the narrowly tuned recurrent excitation function, mutual recurrent inhibition will dominate, leading to a net repulsion of color representations from one another (Felsen, Touryan, & Dan, 2005;Kiyonaga & Egner, 2016), which can serve as a selective partition (figure 6A). (Wei et al, 2012), whereas broadly tuned lateral inhibition facilitates repulsion of distinct bumps of neural activity (Felsen et al, 2005;Kiyonaga & Egner, 2016).…”

A Flexible Model of Working Memory

“…Every neuron i has an associated angle θ i = 2πi/N sensory . Consistent with biological observations [Funahashi et al, 1989;Kiyonaga and Egner, 2016;Kuffler, 1953], connections within a sensory sub-network have a center-surround structure (Fig. 1A, inset).…”

“…Position around the ring corresponds to specific values of an encoded feature, such as orientation or color. Consistent with biological observations [Funahashi et al, 1989;Kiyonaga and Egner, 2016;Kuffler, 1953], connections within a sensory sub-network have a center-surround structure: neurons with similar selectivity share excitatory connections while inhibition is broader (Fig. 1A, inset).…”