Basic operations in working memory: Contributions from functional imaging studies

Abstract

Working memory (WM) constitutes a fundamental aspect of human cognition. It refers to the ability to keep information active for further use, while allowing it to be prioritized, modified and protected from interference. Much research has addressed the storage function of WM, however, its ‘working’ aspect still remains underspecified. Many operations that work on the contents of WM do not appear specific to WM. The present review focuses on those operations that we consider “basic” because they operate in the service of memory itself, by providing its basic functionality of retaining information active, in a stable yet flexible way. Based on current process models of WM we review five strands of research: (1) mnemonic selection of one item amongst others, (2) updating the focus of attention with the selected item, (3) updating the content of visual WM with new item(s), (4) rehearsal of visuospatial information and (5) coping with interference. We discuss the neuronal substrates underlying those operations obtained with functional magnetic resonance imaging and relate them to findings on “executive functions”. The presented data support the view that WM emerges from interactions between higher sensory, attentional and mnemonic functions, with separable neural bases. However, interference processing and the representation of rule switching in WM may demand an extension of the current WM models by executive control functions.

Introduction

Higher-order cognition such as reasoning and problem solving usually requires multiple pieces of information to be related orderly to reach an optimal solution or a correct conclusion. Herein cognition strongly relies on the ability to retain some information active for further use, and to do so in a flexible way allowing information to be prioritized, added or removed. This capability has been named “working memory”, and ever since its introduction by Miller et al. [1], cognitive and clinical neuroscientists have attempted to elucidate its components, associated processes and their underlying neural substrates.

In the neurosciences, much of the research on working memory (WM) has concentrated on one of its most pertinent qualities: to retain information over some period of time, when the physical stimulus is no longer present. To accomplish this, the contents of WM are vastly reduced compared to the wealth of information continuously streamed by our senses. The most commonly noted capacity constraint lies at about 4 items [2]. Also, information about some perceptual details may be lost in the mnemonic stimulus representation. Despite these obvious differences between perception and memory, many processes that operate on sensory stimuli may operate equivalently on WM contents. For example, judging which of two simultaneously presented visual stimuli has a higher contrast can similarly be performed on mental representations. Likewise, attention can be shifted between sensory stimuli or between memory items, and both processes involve largely overlapping brain regions [3], [4].

This notion is in accordance with the contemporary view that WM is not a self-contained system that is composed of a specific set of brain regions separate from those subserving perception [5], [6], [7], [8], [9]. Instead WM is seen as emergent from the interaction of higher sensory, attentional and mnemonic component processes involving their underlying brain regions. For example, higher visual areas subserving the perception of faces and places are also involved in retaining information on faces/places in WM [10]. Activity in perceptual areas is thought to be sustained (or re-activated) by ongoing top-down signals, i.e. attention, from prefrontal and parietal cortex, allowing us to maintain stimulus information in WM for some seconds. In consequence of this concurrence of their structural neuroanatomical bases, perception and memory may also share many operations. Thus, many operations performed on the contents of WM are most likely not uniquely and specifically dedicated to WM. As a consequence, simply listing potential operations on the contents of WM is not appropriate for identifying the basic operations that compose WM (and reviewing the evidence for these). Rather we seek to review those basic operations that make WM emerge – even if they may be “borrowed” from e.g. attentional and general mnemonic systems.

We consider those operations as “basic” that in the context of memory tasks operate in service of memory itself, to provide its basic functionality of retaining information active, in a stable yet flexible way. We focus on operations that go beyond the mere passive maintenance of stimuli, and also leave aside the principles of encoding items into WM. In short, the focus of the present review is on those processes that make WM work, not on all possible operations that can be applied to the contents of WM. This distinction is essential; however, these categories may not be mutually exclusive. We try to define each of the selected operations in the theoretical context provided by Cowan and Oberauer's process model of WM [11], [12] which will be described in some detail below. This will be followed by summaries and discussions of studies on the following operations in WM that we consider as “basic” for its functioning.

(1) Selection: in most circumstances (but only few experiments), not all the representations held in memory are of equal importance all the time, thus requiring some sort of selection or prioritization according to their momentary relevance. For example at an airport, you may have to remember different pieces of information like the terminal number, flight number, gate number and your seat. Their momentary relevance depends on which steps you have already taken. You will not need your seat number before boarding the plane, but by then the terminal number has long become unimportant. The mental selection of one item amongst others as a means of prioritizing items is the first basic operation we review. (2) Focus updating: recently, on the foundation of Cowan's and Oberauer's model [11], [12], we have demonstrated that mnemonic selection and attentional focusing, which naturally co-occur and in consequence had not been studied separately in the studies reviewed in the “selection” section, have dissociable neural substrates. (3) Content updating: returning to the example, arriving at the check-in you may find that your ticket has been re-scheduled to another flight due to overbooking. You therefore have to discard the formerly relevant information and replace it with the newly learned data. In the literature this type of requirement mostly is termed “updating” of memory contents. (4) Rehearsal: in order not to have to look up the flight info repeatedly you may silently rehearse it to prevent forgetting. (5) Coping with interference: when re-scheduled, or just by being confronted with announcements for other flights, your WM contents are easily overwritten or confused with one another. Coping with potentially interfering signals is another basic function that is of vast importance in real-life settings. While selection, updating of the focus of attention and updating the contents of WM provide means for the flexibility of WM, rehearsal and mechanisms of coping with interference are main sources of maintaining items stable in WM.

Following these sections, we summarize and discuss these operations and their potential neural substrates. They are related to “executive functions” and cognitive control.

According to the seminal “multiple-component model” by Baddeley and Hitch [13], WM has been conceptualized as a system made up of two specialized temporary memory buffers (a phonological and a visuospatial store), and a supervisory system (the central executive). While the storage systems hold and refresh memory traces in their dedicated content domain for a few seconds, the central executive is involved in control and regulation of WM. Even though the authors did not explicitly assign its components to structural anatomical regions [14], much of the neuroscientific research stimulated by this model did interpret its components literally and sought for dedicated brain regions. In contrast to Baddeley and Hitch [13], Cowan [11] explicitly considers WM not as a separate system but as a functional state that allows a direct access to information. Specifically, his “embedded-processes model” combines hierarchically arranged faculties comprising long-term memory (LTM), the subset of LTM that is currently activated and the subset of activated memory that is in the focus of attention. The activated memory reflects representations of incoming stimuli and previous cognitive operations. The portion of activated memory that is not in the focus of attention is regarded as passive and hence is prone to decay and interference. In contrast, the items in the focus of attention are maintained actively, but their number is tightly limited. The focus of attention may hold up to four of the activated representations. As commonly assumed for attention, the focus can be driven by exogeneous stimuli or be voluntarily controlled and directed.

Oberauer [12] extended Cowan's model [11] by adding a further component, a more narrow focus of attention that holds only one representation at a time (Fig. 1). The one-element focus and the four-element focus refer to two different functional states of accessibility for cognitive processes. Whereas the four-element focus holds a limited number of representations available for use in ongoing cognitive processes, the one-element attentional focus hosts the single representation that is actually selected as the object of the next cognitive operation. Thus, Oberauer [12] argued that the role of the focus is not to hold a set of memory elements ready for access but to hold a single object already selected for processing.