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

Stokes MG. · 2015

Read 'Activity-silent' working memory in prefrontal cortex: a dynamic coding framework. by Stokes MG. (2015) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4509720.txt.

Introduction

Mark G. Stokes wrote a solo review in Trends in Cognitive Sciences in 2015 on how the brain holds things in mind. Working memory keeps a phone number or a planned move alive for a few seconds. The classic view says cells in the prefrontal cortex, at the front of the brain, keep firing the whole time to hold the trace.

Recordings at the level of cell groups tell a more restless story: activity shifts and can even go quiet while the memory still works. Stokes argues the trace may sit in changing patterns of links between cells, an activity-silent state, rather than in steady firing. His framework aims to reconcile the old delay-firing findings from monkey studies with the newer dynamic picture.

Catalog record

CREATORStokes MG.
YEAR2015
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC4509720

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

‘Activity-silent’ working memory in prefrontal cortex: a dynamic coding framework By Mark G. Stokes Published in: Trends in Cognitive Sciences DOI: 10.1016/j.tics.2015.05.004 ABSTRACT Highlights •WM is thought to depend on persistent maintenance of stationary activity states.•However, population-level analyses reveal that brain activity is highly dynamic.•Accumulating evidence implicates activity-silent neural states for WM.•Dynamic coding suggests that WM is encoded in patterns of functional connectivity. SEARCHING FOR THE WM ENGRAM IN THE PREFRONTAL CORTEX (PFC) There is little doubt that the PFC is important for WM [1]. Some of the earliest studies in nonhuman primates demonstrated that the structural [2] and functional [3] integrity of the PFC is necessary for successful maintenance of information over short delays. Although other brain areas presumably contribute some aspects of WM [4,5], especially modality-specific sensory areas [6,7], the lesion evidence clearly implicates the PFC as a crucial hub necessary for successful maintenance (cf. visual cortex for encoding visual memoranda [8]). This review focuses on neurophysiological recordings in nonhuman primates to disentangle the functional principles of the PFC that mediate WM. Early recordings around the principal sulcus in the monkey PFC discovered an apparent neurophysiological correlate of the WM engram: task-specific delay activity [9,10]. For example, in a delayed reaching task some prefrontal cells increased firing during the delay period, indeed remained active for the duration that response-related information needed to be held in mind, and returned to baseline after the response was executed [9]. A similar sustained firing pattern is observed in a memory-guided saccade (MGS) task [11]. In this now-classic WM task, the subject is presented with a spatial cue at the beginning of the trial that indicates the target for a saccade that can be executed only after a memory delay of a few seconds (Figure 1A). Activity in this memory delay is selective for the specific saccade location, corresponding to the contents of WM (Figure 1B). Prefrontal delay activity is not limited to motor preparation, but also responds selectively to visual memoranda [12]. Similar to the delayed response tasks, PFC activity during visual delay match-to-sample tasks also covaries with the duration of maintenance in WM. Moreover, some cells even maintain their mnemonic role after the presentation of distracting items, indicating a potential correlate of distracter-resistant WM [12]. Evidence for persistent content-specific delay activity has motivated numerous theoretical [1,13,14] and computational [15,16] models based on the idea that representations are kept ‘online’ after stimulus offset in a stable activity state. ‘ACTIVITY-SILENT’ WM However, maintenance in WM is not always accompanied by an unbroken chain of persistent delay activity in the PFC [17,18]. For example, when the duration of the maintenance period is fixed from trial to trial, robust delay activity sometimes emerges only late during the delay period (Figure 1B) [19] (see also [20]). It seems that when the temporal structure of the task can be predicted, neural activity is reserved for the most relevant time point: that is, preparation for the expected memory probe. So-called ramp-up activity implies that delay activity could optimise WM-guided behaviour by preparing for the processing demands of the behavioural response (e.g., temporal orienting of attention [21]). However, the relatively silent moments between encoding and response preparation additionally suggest that the continuity of vigorous delay activity is not always necessary for the continuity of the mental representation [20]. A recent neurophysiological study reported further evidence suggesting that persistent delay activity might not be necessary for accurate WM-guided behaviour [22]. Critically, this study added a dual task demand to the classic MGS task [1]. As in the standard variant, monkeys were required to encode and maintain the location of a saccade target for execution at the end of the trial. However, they were additionally required to attend to a specific spatial location until part way through the WM delay period (Figure 1C). This cognitive manipulation effectively abolished WM-specific delay activity during the dual task period, even on trials in which WM performance was preserved (WM-correct trials). Moreover, when the competing task demands were completed, content-specific delay activity was ‘reawakened’, presumably in time for WM-guided behaviour. Consistent with the evidence from ramp up, these findings suggest that mnemonic delay activity in the PFC is not always critical for maintaining the continuity of WM but can be dynamically re-established when attention is refocused to the task-relevant content. WM does not seem to depend on delay activity but might be maintained in an ‘activity-silent’ format such as functional connectivity. ACTIVITY-SILENT WM STORED IN FUNCTIONAL CONNECTIVITY According to a synaptic model of WM, information can be maintained during such activity-silent periods as a pattern of synaptic weights [23,24], analogous to long-term memory. In this computational model (Figure 2A), activity during encoding temporarily changes synaptic efficacy within the neural network, leaving behind a temporary synaptic memory trace via activity-dependent short-term synaptic plasticity (STSP) [25]. More specifically, calcium kinetics (see Glossary) provide a window of approximately 2 s for STSP in the PFC [23]; however, presumably different time constants could underlie the diversity of mnemonic timescales observed in the PFC [26]. The essential point is that STSP changes the functional connectivity of the network to construct a temporary task-relevant circuit for WM-guided behaviour (see also [27]). According to the original computational model [23], an activity-silent coding scheme is more efficient than persistent firing models. Persistent firing could be energetically expensive, especially if WM is in near-constant use constructing and maintaining an up-to-date world model of the contents of the environment and the rules that govern them. The brain consumes about 20% of total energy and it is estimated that action potentials and postsynaptic effects of glutamate account for much of this expenditure (47% and 34%, respectively [28]). Therefore, neural economy places important constraints on ecologically optimal coding strategies [29]. For example, it has been estimated that less than 1% of neurons can be activated simultaneously [30]. It is difficult to rule out energy-efficient variants of persistent activity models of WM (e.g., WM codes could be sparsified across a cell population [31]). Nevertheless, a coding format that does not depend on an unbroken chain for persistent firing would have a clear cost advantage (i.e., sparsification in time). WM coded in patterns of functional connectivity might also provide a neurobiologically plausible mechanistic account for WM-guided behaviour as a generalised state-dependent processing. Memories are not stored as an active representation but rather change the functional architecture of the neural network for future processing. As in long-term memory, WM is expressed by the way the network responds to new input. State-dependent readout would avoid the need for explicit mechanisms for comparing the internal representation with a separate representation reflecting the new input [23,32]. Moreover, because memories are stored in a format that is qualitatively different from perceptual representations, their informational content could be more resistant to interference from and/or confusion with other activity-based representations [33]. Despite considerable theoretical appeal, the synaptic model of WM is relatively difficult to test empirically. Extensive evidence confirms the prevalence of STSP [25] but its functional role has not yet been fully established. Microelectrode neurophysiology can only indirectly infer connectivity via interactions between simultaneous recordings. The gold-standard evidence for a monosynaptic connection is correlated spiking between neurons. In practice, however, the probability of sampling any

Read the complete text