Trisecting representational states in short-term memory

Abstract

The ability to hold information briefly in mind in the absence of external stimulation forms the core of much of higher-order cognition. This ability is referred to as short-term memory (STM). However, single-term labels such as this belie the complexity of the underlying construct. Here, we review evidence that STM is an amalgamation of three qualitatively distinct states. We argue that these distinct states emerge from the combination of frontal selection mechanisms (often considered the domain of attention and cognitive control), medial temporal binding mechanisms (often considered the domain of long-term memory, LTM), and synaptic plasticity. These various contributions lead to a single representation amenable to elaborated processing (focus of attention), a limited set of active representations among which attention can be flexibly switched (direct-access region), and passive representations whose residual traces facilitate re-activation (activated LTM). We suggest that selection and binding mechanisms are typically engaged simultaneously, providing multiple forms and routes of short-term maintenance. We propose that such a framework can resolve discrepancies among recent studies that have attempted to understand the relationship between attention and STM on the one hand, and between LTM and STM on the other. We anticipate that recent advances in neuroimaging and neurophysiology will elucidate the mechanisms underlying shifts and transformations among these representational states, providing a window into the dynamic processes of higher-order cognition.

Keywords: attention; long-term memory; medial temporal lobe; prefrontal cortex; working memory.

FIGURE 1

Three-state model of memory. Left: illustration of the task used by Oberauer (2002) to test the three-state model of memory. Participants hold in mind two sets of digits. Red frames indicate the active set whose digits are candidates for processing (e.g., top set). Black frames indicate the passive set, which is recalled at the end of the trial, but not the subject of operations. Mathematical operations are applied to the active set thereby updating its contents. In this example, “-4” is applied to “5,” resulting in “1.” Subsequently, “+2” is applied to the result, yielding “3” and so on. At the end of the trial, all digits are recalled. In this example, going left to right starting from the top, recall would be “6,” “3,” “9,” “7,” “8,” “2.” Right: depiction of the representational states of STM according to the model. The modeled scenario reflects the moment at which “-4” is presented. The cue “-4” draws the focus of attention (FA) to its corresponding frame. The number associated with that frame, “5,” is recalled through location-digit bindings. All location-digit bindings for the active set are maintained through the direct-access region (DAR). It is assumed that items are also inter-associated with each other, as well as other items that are not actively maintained (e.g., passive set). The passive set, which is not contextually bound, is held in activated long-term memory (aLTM).

FIGURE 2

Assessing representational states via serial positions. Illustration of the task used by Nee and Jonides (2008) to dissociate the focus of attention from the direct-access region. Three words are sequentially presented followed by a mask and a recognition probe. Timing information depicts the onset of each stimulus relative to the start of the trial. As each memorandum is presented, it is assumed that it becomes the focus of attention. Thus, upon presentation of “TOOL,” “TOOL” is the focus of attention. When “LAKE” is presented, the focus of attention switches to “LAKE.” As each item is presented, it is bound to the temporal context of the trial. At the end of encoding, the last word, “DIRT,” is the focus of attention. Two kinds of probes are depicted. The left depicts a scenario where “LAKE” is the probe. This activates its corresponding representation. The bindings to the trial context, maintained by the direct-access region, verify that “LAKE” is an old probe resulting in a match decision. This elicits activation in the medial temporal lobe (MTL). The right depicts an alternative scenario where “DIRT” is the probe. Once again, “DIRT” activates its corresponding representation. In this case, however, “DIRT” is the focus of attention, so it can be verified immediately without the need to retrieve contextual information. This elicits activation in inferior temporal cortex (ITC) and ventral parietal cortex (VPC). Activation data adapted from Nee and Jonides (2011).

FIGURE 3

Neural evidence for a three-state model of memory. (A) The task used by Nee and Jonides (2013) to examine neural correlates of the focus of attention (FA), direct-access region (DAR), and aLTM in visual STM. The task involved sequential presentation of five faces followed by a mask and a recognition probe. In this task, the number of presented items exceeded the capacity of the direct-access region. So, it was assumed that the least recent items (i.e., first presented items) would no longer be contextually bound and be represented in aLTM rather than the direct-access region. Reprinted from Nee and Jonides (2013) with permission from Elsevier. (B) Conjunction of results across a six-word version of the paradigm depicted in Figure 2 (Nee and Jonides, 2011) and the five-face paradigm depicted in (A) (Nee and Jonides, 2013). Across both studies, probes matching the focus of attention activated ventral posterior parietal cortex (VPC), probes matching the direct-access region activated the medial temporal lobe (MTL), and probes matching the aLTM activated the ventrolateral prefrontal cortex (VLPFC). In all cases, activations related to a given state were dissociable from those involved in the access of other representational states. This triple dissociation supports the three-state model of memory. Reprinted from Nee and Jonides (2013) with permission from Elsevier.

FIGURE 4

Neural three-state model of memory. Top: a hypothetical task requiring the maintenance of visual objects and attention shifting among them. Four objects are presented and encoded into STM. Following a retention interval, a cue directs the focus of attention to one of the objects. A recognition probe is presented in the cued location requiring a match/non-match decision. This is followed by another retention interval, a second cue, and a second probe. Bottom: a model demonstrating relevant areas of the brain and hypothesized psychological and neural processes. The model has been simplified for depictive purposes. For full details, consult the text. Mechanisms associated with the focus of attention are depicted in green while mechanisms associated with the direct-access region are depicted in red. Object information is presumed to be represented in inferior temporal cortex (ITC) while spatial information is presumed to be represented in the IPS. Each of these posterior areas is connected to a corresponding frontal area. The IPS is connected to the SFS, a region that is commonly referred to as the frontal eye fields. The ITC is connected to the ventrolateral prefrontal cortex (VLPFC). Each of these frontal areas selects information represented in respective posterior areas. Finally, the medial temporal lobe (MTL) is connected to both the IPS and ITC and synchronizes their activity. (A) During the retention interval, the MTL synchronizes the activity of the IPS and ITC in the theta range. Within each theta cycle, stimulus-specific neurons fire in the gamma range. Thus, gamma activity nested within theta activity reflects the cycling of items within a set. Moreover, item-location bindings are implemented by MTL-mediated synchronized activity in the ITC and IPS. The connections and synchrony correspond to the direct-access region. At the same time, the VLPFC acts upon the ITC to support item-based maintenance. Here also, individual items are nested within oscillatory activity. These mechanisms periodically maintain the activity corresponding to each item in STM and correspond to the focus of attention cycling among the direct-access region. The SFS performs a similar function in concert with the IPS to cycle among locations. (B) When a cue directs attention to a spatial location, cycling ceases. Instead, the SFS fixates on the cued location thereby forming the focus of attention. Sustained spatial attention then activates the corresponding object through connections established by previous synchrony. While attention is sustained, only the attended location and corresponding object are instantiated by active neural firing. The synchronous activity between neurons corresponding to location representation and object representation strengthen the bindings between them (thickened red line) potentiating future cued retrieval.