Between persistently active and activity-silent frameworks: novel vistas on the cellular basis of working memory

Abstract

Recent work has revealed important new discoveries on the cellular mechanisms of working memory (WM). These findings have motivated several seemingly conflicting theories on the mechanisms of short-term memory maintenance. Here, we summarize the key insights gained from these new experiments and critically evaluate them in light of three hypotheses: classical persistent activity, activity-silent, and dynamic coding. The experiments discussed include the first direct demonstration of persistently active neurons in the human medial temporal lobe that form static attractors with relevance to WM, single-neuron recordings in the macaque prefrontal cortex that show evidence for both persistent and more dynamic types of WM representations, and noninvasive neuroimaging in humans that argues for activity-silent representations. A key insight that emerges from these new results is that there are several neural mechanisms that support the maintenance of information in WM. Finally, based on established cognitive theories of WM, we propose a coherent model that encompasses these seemingly contradictory results. We propose that the three neuronal mechanisms of persistent activity, activity-silent, and dynamic coding map well onto the cognitive levels of information processing (within focus of attention, activated long-term memory, and central executive) that Cowan's WM model proposes.

Keywords: attractors; dynamic coding; persistent activity; single-neuron recordings; static coding; working memory.

Figure 1

Summary of theories of the neuronal mechanisms supporting working memory. (A) The persistent activity framework proposes that memoranda are maintained by the sustained firing of stimulus‐specific groups of neurons. A decoder trained in one period of time should be able to decode information at a different point of time (“1” versus “2” periods indicated). (B) The activity‐silent framework proposes that information held in WM is not visible by observing the activity of individual neurons. (C) The dynamic coding framework proposes that the neurons carrying information about a specific item change as function of time relative to the onset of the maintenance period. For example, some neurons encode the identity of an item only at a specific period of time. In the figure, three neurons are shown, all of which represent item A, but during different periods of time, with some neurons “ramping down” their activity (top), whereas others firing only during specific periods of time. A decoder trained at one point of time will thus not generalize to a different point of time (“1” versus “2” periods indicated).

Figure 2

Persistent activity represented by attractor dynamic. (A) Example of a persistently active concept cell recorded in the human amygdala. Subjects memorized up to three images presented sequentially (encoding 1–3). Top: post stimulus time histogram. Middle: periods of significance (black) between the preferred versus nonpreferred stimuli of this cell. Bottom: raster plot of trials reordered according to condition. During maintenance, the activity of this cell is characterized by sustained activity only when the preferred image of the cell is held in memory (blue) but not when other stimuli are held in memory (gray). Adapted from Ref. 17. (B) Trajectories in neural state space formed by a population of persistently active concept cells in the human MTL. Trajectories are projected into the 3D space formed by the three demixed principal components (dPCs) associated with picture identity. Periods of time shown are encoding (thin line) and maintenance (thick line). Colors mark different images. Note how during maintenance, activity settles at points in space that separate by memory content (attractors). Adapted from Ref. 17. (C) Neuron whose activity is indicative of a continuous attractor during a delayed oculomotor task in rhesus monkey. Firing rate of stimulus‐selective neurons is sorted vertically according to preferred location. Note how activity drifts away from the initial position (left) as time progresses. This drift predicts behavioral errors (right). Adapted from Ref. 50. (D) Persistent activity recorded in mouse ALM during a task with variable delay durations. Blue color marks prefered location. (E) Population activity in mouse ALM shows characteristic of a discrete attractor. Shown is a projection of the population activity onto the axis that maximally distinguishing between the two possible conditions (left or right). Blue color (left panel) denotes correct lick‐right trials, and red denotes correct lick‐left trials. Dark blue and dark red (right panel) denote incorrect lick‐right and lick‐left trials, respectively. Dashed lines denote trajectories of unperturbed correct trials, whereas solid lines denote perturbed trials. Light blue band on the top shows time of photoinhibition. Note how the neural activity after offset of inhibition is pulled toward one of the two possible trajectories, as expected from an attractor. Adapted from Ref. 7.

Figure 3

Different possible mechanisms of coding information in working memory. (A) Task design of retro‐cue paradigm. A trial starts with presentation of two items from two different categories. After a delay period, a retro‐cue indicates which of the two is relevant for the probe that follows (here: face). After a delay, a probe is shown that requires the subject to perform the task indicated by the preceding retro‐cue. Afterward, a second retro‐cue indicates which stimulus is relevant for the second probe, thereby leading to reactivation of items in WM. Adapted from Ref. 40. (B) Decoding of the identity of the maintained category based on the BOLD‐fMRI signal in the experiment illustrated in A. Circles, triangles, and squares mark the presentation of stimuli, retro‐cue, and probe, respectively. After presentation of the first retro‐cue, decoding accuracy for the unattended item drops to chance but returns after the second retro‐cue when subjects are instructed to bring the back into the focus of attention (blue line). Adapted from Ref. 40. (C) Dynamic coding of information in mice posterior parietal cortex measured using calcium imaging. Note the sequential activation of cells encoding spatial information. Color code indicates fluorescence intensity. Adapted from Ref. 52. (D) Mix of stable and dynamic coding in a task in which reward cues are shown as distractors during WM maintenance. Each (x, y) point shows the performance of a decoder trained and tested at a different point of time. The square‐block “red” shows across‐time generalization of the decoder during the first maintenance period, followed by dynamic coding (diagonal red in lower right). Adapted from Ref. 56. (E) Stable mnemonic subspace coexists with dynamic activity. Stimulus PC1 and stimulus PC2 represent the mnemonic subspace, whereas time PC1 represents the dynamic component. Colors mark locations on the screen that animals needed to memorize. Adapted from Ref. 54.

Figure 4

Illustration of Cowan's model of working memory with possible neuronal mechanism associated with each component indicated. Under this hypothesis, information currently in the focus of attention is represented by classical persistent activity. Result from the retro‐cue paradigm suggests that information outside of the focus of attention is maintained by activity‐silent states, a form of coding compatible with the activated long‐term memory (LTM) part of the model. Finally, the central executive part of the model performs different functions as a trial progresses, thereby implementing task demands. Here, we propose that dynamic coding is a reflection of this process.