Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
Review
. 2015 Jul;19(7):394-405.
doi: 10.1016/j.tics.2015.05.004. Epub 2015 Jun 4.

'Activity-silent' working memory in prefrontal cortex: a dynamic coding framework

Mark G Stokes  1
Affiliations
Review

'Activity-silent' working memory in prefrontal cortex: a dynamic coding framework

Mark G Stokes. Trends Cogn Sci. 2015 Jul.

Abstract

Working memory (WM) provides the functional backbone to high-level cognition. Maintenance in WM is often assumed to depend on the stationary persistence of neural activity patterns that represent memory content. However, accumulating evidence suggests that persistent delay activity does not always accompany WM maintenance but instead seems to wax and wane as a function of the current task relevance of memoranda. Furthermore, new methods for measuring and analysing population-level patterns show that activity states are highly dynamic. At first glance, these dynamics seem at odds with the very nature of WM. How can we keep a stable thought in mind while brain activity is constantly changing? This review considers how neural dynamics might be functionally important for WM maintenance.

PubMed Disclaimer

Figures

Figure 1
Figure 1
Working memory (WM) delay activity varies with task relevance. (A) Schematic of the influential memory-guided saccade (MGS) task (e.g., [1]). On each trial, the subject is presented with a memory location and, after a delay period, executes a MGS. (B) Neurophysiological recordings from monkey prefrontal cortex [near the principal sulcus (PS)] during the delay period reveal content-specific activity (location-specific activity is shown in black, population mean activity is shown in grey for reference). However, content-specific delay activity also varies with the current task relevance of the memory item, resulting in a ‘ramp up’ in anticipation of the response. Adapted from , with permission from Oxford University Press. (C) A more recent dual-task experiment partially overlaid an attention task and the standard MGS design. Although initial encoding of the saccade location was robust (red trace, mean activity shown in grey), content-specific activity was effectively abolished by the dual task during the memory delay (blue trace, from −1000 to onset of the attention target). Critically, location-specific information was ‘reactivated’ at the end of the dual task (blue trace, after onset of the attention target), presumably reflecting a shift in task focus to the MGS. Adapted from with permission from Nature Publishing Group. Such evidence suggests that delay activity reflects the task relevance of memoranda. Gaps in content-specific activity further suggest that WM could be maintained in an ‘activity-silent’ neural state.
Figure 2
Figure 2
Maintaining ‘activity-silent’ working memory (WM) in functional connectivity. (A) Schematic of the synaptic model of WM described in . Task-relevant input (left-side horizontal arrows, blue for ‘Memory A’ and red for ‘Memory B’) drives a stimulus-specific activity state (filled circles) that in turn triggers a specific pattern of short-term synaptic plasticity between cells (bold arrows). Memory is read out from this synaptic trace via the context-dependent response at retrieval (black filled circles, same for ‘Memory A’ and ‘Memory B’). The probe-driven response will be patterned by the hidden state of synaptic efficacy, resulting in a discriminable output pattern (right-side horizontal arrows). (B) Empirical evidence for content-specific functional networks. Simultaneous recordings in rat frontal cortex revealed direction-specific patterns of synaptic efficacy (red arrows) between cells [putative pyramidal (triangles); putative interneuron (circle); unclassified (square)], encoding direction during a WM-based maze task (left panel; adapted from with permission from Nature Publishing Group). This is consistent with a role for short-term synaptic plasticity in WM. In the monkey prefrontal cortex (PFC) (PS, principal sulcus), different task rules are associated with specific functional networks (synchronised electrodes for rule ‘A’ in blue, rule ‘B’ in red) coupled by synchrony at ∼30 Hz (left panel; adapted from [36]). This is consistent with the idea that coherence could also play a role in constructing functional networks for flexible behaviour.
Figure 3
Figure 3
Representing dynamic population coding. (A) Schematic illustration of three different temporal coding profiles in three hypothetical cells: transient, sustained, and dynamic. Colours correspond to different experimental conditions (e.g., memoranda). (B) These dynamics can be visualised in a 3D state-space plot (for example, activity in three separate neurons). The trajectory through state space illustrates the dynamic profile. The trajectory for transient coding is illustrated as a simple path to a specific point in state space with immediate return. The sustained profile is also shown as a simple return trajectory, but with additional time spent at the target point in state space. Finally, dynamic coding is shown as a more complex path through state space. (C) The same dynamics can be computed for n dimensions (where n is the size of the neural population) and visualised using a 2D representation of cross-temporal pattern analysis. Essentially, this approach measures the consistency of a population code between time points (i.e., Time 1 along the y-axis, Time 2 along the x-axis). In this schematic, pattern consistency is colour coded red whereas nonconsistent patterns are colour coded green. Upper left panel illustrates the basic structure: diagonal values reflect time-specific patterns [p(t1,t1), p(t2,t2)]; off-diagonals reflect cross-temporal generalisation [p(t1,t2)]. Upper right panel illustrates the expected profile for transient coding, with reliable patterns at p(t1,t1) but not p(t2,t2). In the absence of coding at t2, it also follows that there is no cross-time generalisation at p(t1,t2). Lower left panel illustrates sustained coding, with consistent patterns at p(t1,t1) and p(t2,t2) but critically also significant cross-temporal generalisation p(t1,t2). This indicates a shared coding pattern at t1 and t2. By contrast, the lower right plot illustrates dynamic coding: consistent patterns at p(t1,t1) and p(t2,t2) but critically no significant cross-temporal generalisation [p(t1,t2)] (see also [39]).
Figure 4
Figure 4
Evidence for dynamic coding of working memory (WM) in monkey prefrontal cortex (PFC). (A) The population-level geometric measures described in Figure 3 were applied in to track WM coding in the monkey PFC. A state-space representation based on the dimensionality-reduced population response illustrates dynamic trajectories for each memory condition (three conditions, colour coded). (B) These dynamics can be seen clearly in the 2D cross-temporal analysis matrix, with robust WM-specific activity along the within-time diagonal axis and relatively poor temporal cross-generalisation at off-diagonal time points. Adapted from .
Figure 5
Figure 5
Context-dependent readout for ‘activity-silent’ working memory (WM). (A) Synaptic WM is effectivity activity silent; however, the information content can be inferred from a shift in the behaviour of the network. This can be probed using an impulse of activity to drive the network (Box 1). (B) We previously observed evidence for impulse-driven reactivation of WM in the prefrontal cortex. Using multivariate statistics, we found that a neutral stimulus triggered a distinct pattern of activity that reflected the content of WM (blue trace) although the driving stimulus was exactly the same for all WM conditions (from [44]). (C) A temporary shift in the response profile of the network could also provide a basis for context-dependent processing during WM-guided behaviour. For example, we found that during WM-guided behaviour population-level activity states quickly evolved from representing the physical properties of the choice stimuli (at 125 ms; stimuli are coloured coded) to the decision-relevant coding from ∼150 ms (i.e., target versus non-target; from [44]). (D) This process can be schematised as a context-dependent path through activity state space. A functional shift in the hidden state effectively conditions the context-dependent input/output behaviour of the network to map stimulus-specific activity states to the context-relevant position in state space (upper panel; from [44]). This flexible mapping could be framed as a context-dependent decision process. In the lower panel, the results from (C) are replotted as the accumulation of evidence for the ‘go’ or ‘no-go’ response as a function of the current rules (i.e., if context 1, then evidence for stimulus 1 supports a ‘go’ response, but stimuli 2 and 3 provide counter-evidence; upper panel adapted from [44]).
Figure 6
Figure 6
Summary schematic of dynamic coding for working memory (WM)-dependent behaviour. The initial input triggers a specific pattern in activity state that in turn alters the underlying hidden state of the network via a temporary shift in functional connectivity (e.g., short-term synaptic plasticity , or coherence [35]). Activity-dependent changes in hidden state drive a dynamic trajectory during the initial high-energy phase via the reciprocal interaction between hidden states and the activity states that modulate them . After activity has relaxed to baseline levels, the hidden state remains patterned according to the WM item. Although in principle this temporary hidden state could be ‘activity silent’, any spontaneous activity in the network will be patterned according to the WM context, resulting in a WM-specific activity state during spontaneous firing . This kind of ‘baseline emission’ could help explain the content-specific delay activity observed under some circumstances . Increasing the level of network activity via attention/rehearsal mechanisms could increase the discriminability of the activity state, resulting in ‘ramp-up’ delay activity or task-dependent ‘reactivations’ . Finally, when the critical memory cue is presented, the context-dependent response maps activity states for WM-guided behaviour (e.g., match/non-match decision; see [44]).
Figure I
Figure I
Schematic of layered neural states. The neural state comprises the ‘activity state’ that is measured in typical experiments, but also ‘hidden states’ such as functional connectivity and structural connectivity. Short-term changes in these hidden states could play an important role in high-level cognition, including working memory.
See this image and copyright information in PMC

References

    1. Goldman-Rakic P.S. Cellular basis of working memory. Neuron. 1995;14:477–485. - PubMed
    1. Butters N., Pandya D. Retention of delayed-alternation: effect of selective lesions of sulcus principalis. Science. 1969;165:1271–1273. - PubMed
    1. Stamm J.S. Electrical stimulation of monkeys’ prefrontal cortex during delayed-response performance. J. Comp. Physiol. Psychol. 1969;67:535–546. - PubMed
    1. Pasternak T., Greenlee M.W. Working memory in primate sensory systems. Nat. Rev. Neurosci. 2005;6:97–107. - PubMed
    1. Postle B.R. Neural bases of the short-term retention of visual information. In: Jolicoeur P., editor. Mechanisms of Sensory Working Memory: Attention and Performance XXV. AcademicPress; 2015. pp. 43–58.

Publication types

LinkOut - more resources