Flexible cognitive resources: competitive content maps for attention and memory

Abstract

The brain has finite processing resources so that, as tasks become harder, performance degrades. Where do the limits on these resources come from? We focus on a variety of capacity-limited buffers related to attention, recognition, and memory that we claim have a two-dimensional 'map' architecture, where individual items compete for cortical real estate. This competitive format leads to capacity limits that are flexible, set by the nature of the content and their locations within an anatomically delimited space. We contrast this format with the standard 'slot' architecture and its fixed capacity. Using visual spatial attention and visual short-term memory as case studies, we suggest that competitive maps are a concrete and plausible architecture that limits cognitive capacity across many domains.

Box 1

Examples of cortical maps

Figure 1

At top, a competitive map representation within a bounded, two-dimensional space. (a) This panel depicts how each location represents value in the space, with items represented by peaks of activation. Each peak also suppresses surrounding space, inhibiting nearby competitors [8]. This surround suppression will limit the number of items that can be simultaneously maintained. If the item spacing is dense, as in (b), the space will be inefficiently used. If the spacing is sparse, as in (c), the space is efficiently used and capacity is maximized, though still limited (by approximately six items in this case). (d) This panel shows that anatomical boundaries (e.g., the visual hemifield divisions of V1) can mimic spacing effects by eliminating mutual inhibition. At bottom, a slot representation limited to four independent items.

Figure 2

(a) Architecture of spatial attention (adapted from [19]). A network of areas form a competitive target map that subserves spatial attention, as well as eye movements. Peaks of activity specify retinotopic coordinates of feature data in earlier visual cortices which are shown, highly simplified, as a stack of aligned areas divided into right and left hemifields with the fovea in the center. In object recognition areas, cells have large receptive fields shown here as a heavy black outline for the receptive field of one cell that specializes in identifying corkscrews. These cells must rely on attention to bias input in favor of the target and suppress surrounding distractors, so that only a single item falls in the receptive field at any one time. The surround suppression must be imposed in early retinotopic areas, because the large fields in object recognition cannot locally modulate sensitivity. (b) Resource limits in multiple object tracking (MOT) tasks. In MOT, a participant is asked to track multiple moving objects (marked here in red for illustration only) among visually identical distractors, which requires constant spatial selection of those objects. When concurrent MOT displays are arranged within visual quadrants, tracking within two vertically arranged displays leads to ‘resource drains’, where performance drops. However, when arranged horizontally, resources are ‘independent’, because performance is virtually unaffected [33]. At bottom, the flexible map account predicts this effect, because the visual hemifield boundary strongly blocks inhibition horizontally, but only weakly blocks it vertically [28]. A strong competitive map account of such effects predicts that almost all performance limitations in this task can be ascribed to competition within a spatial map representing target positions [10].

Figure 3

Maps as limits on visual memory capacity. At left, data from [38], showing that VSTM can hold fewer objects as they become more complex, with an ultimate limit of approximately four total objects that can be held. The remaining boxes present a series of maps that might create visual short-term memory limitations. The first set shows two types of visual displays, complex shapes and simple letters. The next column depicts spatial selection maps (though feature selection, e.g., by color, is equally likely). In this case, one or two locations are selected, biasing competition within multiple hierarchical levels of visual data maps (V1–V4, MT, IT) relevant to recognition of those objects. Critically, because the space for complex shapes is more densely packed and/or requires simultaneous activation of more locations to encode the complex shape information, few shapes can be represented concurrently without their representations degrading, and therefore only one shape should be attended at once in the selection maps. Letters, by contrast, have a well-spaced map of high-level identities created by vast experience, and, therefore, multiple letter identities can be reliably encoded at once, allowing multiple locations to be attended at once in the selection maps. On the right, a hypothetical ‘spatial’ memory map that holds pointers to previously seen visual data (or more likely, pointers to selection maps that point to those data). These could be subserved by connections between activation maintenance structures in frontal cortex and parietal selection maps.