Restoration of fMRI Decodability Does Not Imply Latent Working Memory States

Are latent working memory items retrieved from long-term memory?

Switching one's focus of attention between to-be-remembered items in working memory (WM) is critical for cognition, but the mechanisms by which this is accomplished are unclear. A long-term memory (LTM) account suggests that switching attention away from an item, and passively retaining and reactivating such "latent" items back into the focus of attention involves episodic LTM retrieval processes, even for delays of only a few seconds. We tested this hypothesis using a two-item, double-retrocue WM task that requires participants to switch attention away from and reactivate items followed by subsequent LTM tests for reactivated items from the initial WM task (vs. continuously retained or untested control items). We compared performance on these tests between older adults (a population with LTM deficits) and young adults with either full (Experiment 1) or divided (Experiment 2) attention during the WM delay periods. The effects of reactivating latent items, as well as ageing and divided attention, had significant effects on WM performance, but did not interact with or systematically affect subsequent LTM for reactivated versus control items on item-, location-, or associative-recognition memory judgements made with either high or low confidence. Experiment 3 confirmed that these effects did not depend on whether or not young participants were warned about the subsequent LTM tests before performing the WM task. These dissociations between WM and LTM are inconsistent with the LTM account of latent WM; they are more consistent with the dynamic processing model of WM (Current Directions in Psychological Science).

The computational foundations of dynamic coding in working memory

Working memory (WM) is a fundamental aspect of cognition. WM maintenance is classically thought to rely on stable patterns of neural activities. However, recent evidence shows that neural population activities during WM maintenance undergo dynamic variations before settling into a stable pattern. Although this has been difficult to explain theoretically, neural network models optimized for WM typically also exhibit such dynamics. Here, we examine stable versus dynamic coding in neural data, classical models, and task-optimized networks. We review principled mathematical reasons for why classical models do not, while task-optimized models naturally do exhibit dynamic coding. We suggest an update to our understanding of WM maintenance, in which dynamic coding is a fundamental computational feature rather than an epiphenomenon.

Representation and computation in visual working memory

The ability to sustain internal representations of the sensory environment beyond immediate perception is a fundamental requirement of cognitive processing. In recent years, debates regarding the capacity and fidelity of the working memory (WM) system have advanced our understanding of the nature of these representations. In particular, there is growing recognition that WM representations are not merely imperfect copies of a perceived object or event. New experimental tools have revealed that observers possess richer information about the uncertainty in their memories and take advantage of environmental regularities to use limited memory resources optimally. Meanwhile, computational models of visuospatial WM formulated at different levels of implementation have converged on common principles relating capacity to variability and uncertainty. Here we review recent research on human WM from a computational perspective, including the neural mechanisms that support it.

Pinging the brain to reveal the hidden attentional priority map using encephalography

Attention has been usefully thought of as organized in priority maps - putative maps of space where attentional priority is weighted across spatial regions in a winner-take-all competition for attentional deployment. Recent work has highlighted the influence of past experiences on the weighting of spatial priority - called selection history. Aside from being distinct from more well-studied, top-down forms of attentional enhancement, little is known about the neural substrates of history-mediated attentional priority. Using a task known to induce statistical learning of target distributions, in an EEG study we demonstrate that this otherwise invisible, latent attentional priority map can be visualized during the intertrial period using a 'pinging' technique in conjunction with multivariate pattern analyses. Our findings not only offer a method of visualizing the history-mediated attentional priority map, but also shed light on the underlying mechanisms allowing our past experiences to influence future behavior.

Are irrelevant items actively deleted from visual working memory?: No evidence from repulsion and attraction effects in dual-retrocue tasks

Some theories propose that working memory (WM) involves the active deletion of irrelevant information, including items that were retained in WM, but are no longer relevant for ongoing cognition. Considerable evidence suggests that active-deletion occurs for categorical representations, but whether it also occurs for recall of features that are typically bound together in an object, such as line orientations, is unclear. In two experiments, with or without binding instructions, healthy young adults maintained two orientations, focused attention to recall the orientation cued first, and then switched attention to recall the orientation cued second, at which point the uncued orientation was no longer relevant on the trial. In contrast to the active-deletion hypothesis, the results showed that the no-longer-relevant items exerted the strongest bias on participants' recall, which was either repulsive or attractive depending on both the degree of difference between the target and nontarget orientations and the proximity to cardinal axes. We suggest that visual WM can bind features like line orientations into chunked representations, and an irrelevant feature of a chunked object cannot be actively deleted - it biases recall of the target feature. Models of WM need to be updated to explain this and related dynamic phenomena.

Precision of phonological errors in aphasia supports resource models of phonological working memory in language production

Working memory (WM) is critical for many cognitive functions including language production. A key feature of WM is its capacity limitation. Two models have been proposed to account for such capacity limitation: slot models and resource models. In recent years, resource models have found support in both visual and auditory perception, but do they also extend to production? We investigate this by analyzing sublexical errors from four individuals with aphasia. Using tools from computational linguistics, we first define the concept of "precision" of sublexical errors. We then demonstrate that such precision decreases with increased working memory load, i.e., word length, as predicted by resource models. Finally, we rule out alternative accounts of this effect, such as articulatory simplification. These data provide the first evidence for the applicability of the resource model to production and further point to the generalizability of this account as a model of resource division in WM.

Priority-based transformations of stimulus representation in visual working memory

How does the brain prioritize among the contents of working memory (WM) to appropriately guide behavior? Previous work, employing inverted encoding modeling (IEM) of electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) datasets, has shown that unprioritized memory items (UMI) are actively represented in the brain, but in a “flipped”, or opposite, format compared to prioritized memory items (PMI). To acquire independent evidence for such a priority-based representational transformation, and to explore underlying mechanisms, we trained recurrent neural networks (RNNs) with a long short-term memory (LSTM) architecture to perform a 2-back WM task. Visualization of LSTM hidden layer activity using Principal Component Analysis (PCA) confirmed that stimulus representations undergo a representational transformation–consistent with a flip—while transitioning from the functional status of UMI to PMI. Demixed (d)PCA of the same data identified two representational trajectories, one each within a UMI subspace and a PMI subspace, both undergoing a reversal of stimulus coding axes. dPCA of data from an EEG dataset also provided evidence for priority-based transformations of the representational code, albeit with some differences. This type of transformation could allow for retention of unprioritized information in WM while preventing it from interfering with concurrent behavior. The results from this initial exploration suggest that the algorithmic details of how this transformation is carried out by RNNs, versus by the human brain, may differ.

Pinging the brain with visual impulses reveals electrically active, not activity-silent, working memories

Persistently active neurons during mnemonic periods have been regarded as the mechanism underlying working memory maintenance. Alternatively, neuronal networks could instead store memories in fast synaptic changes, thus avoiding the biological cost of maintaining an active code through persistent neuronal firing. Such "activity-silent" codes have been proposed for specific conditions in which memories are maintained in a nonprioritized state, as for unattended but still relevant short-term memories. A hallmark of this "activity-silent" code is that these memories can be reactivated from silent, synaptic traces. Evidence for "activity-silent" working memory storage has come from human electroencephalography (EEG), in particular from the emergence of decodability (EEG reactivations) induced by visual impulses (termed pinging) during otherwise "silent" periods. Here, we reanalyze EEG data from such pinging studies. We find that the originally reported absence of memory decoding reflects weak statistical power, as decoding is possible based on more powered analyses or reanalysis using alpha power instead of raw voltage. This reveals that visual pinging EEG "reactivations" occur in the presence of an electrically active, not silent, code for unattended memories in these data. This crucial change in the evidence provided by this dataset prompts a reinterpretation of the mechanisms of EEG reactivations. We provide 2 possible explanations backed by computational models, and we discuss the relationship with TMS-induced EEG reactivations.

A dynamic neural field model of continuous input integration

The ability of neural systems to turn transient inputs into persistent changes in activity is thought to be a fundamental requirement for higher cognitive functions. In continuous attractor networks frequently used to model working memory or decision making tasks, the persistent activity settles to a stable pattern with the stereotyped shape of a "bump" independent of integration time or input strength. Here, we investigate a new bump attractor model in which the bump width and amplitude not only reflect qualitative and quantitative characteristics of a preceding input but also the continuous integration of evidence over longer timescales. The model is formalized by two coupled dynamic field equations of Amari-type which combine recurrent interactions mediated by a Mexican-hat connectivity with local feedback mechanisms that balance excitation and inhibition. We analyze the existence, stability and bifurcation structure of single and multi-bump solutions and discuss the relevance of their input dependence to modeling cognitive functions. We then systematically compare the pattern formation process of the two-field model with the classical Amari model. The results reveal that the balanced local feedback mechanisms facilitate the encoding and maintenance of multi-item memories. The existence of stable subthreshold bumps suggests that different to the Amari model, the suppression effect of neighboring bumps in the range of lateral competition may not lead to a complete loss of information. Moreover, bumps with larger amplitude are less vulnerable to noise-induced drifts and distance-dependent interaction effects resulting in more faithful memory representations over time.

Persistent Activity During Working Memory From Front to Back

Working memory (WM) extends the duration over which information is available for processing. Given its importance in supporting a wide-array of high level cognitive abilities, uncovering the neural mechanisms that underlie WM has been a primary goal of neuroscience research over the past century. Here, we critically review what we consider the two major "arcs" of inquiry, with a specific focus on findings that were theoretically transformative. For the first arc, we briefly review classic studies that led to the canonical WM theory that cast the prefrontal cortex (PFC) as a central player utilizing persistent activity of neurons as a mechanism for memory storage. We then consider recent challenges to the theory regarding the role of persistent neural activity. The second arc, which evolved over the last decade, stemmed from sophisticated computational neuroimaging approaches enabling researchers to decode the contents of WM from the patterns of neural activity in many parts of the brain including early visual cortex. We summarize key findings from these studies, their implications for WM theory, and finally the challenges these findings pose. Our goal in doing so is to identify barriers to developing a comprehensive theory of WM that will require a unification of these two "arcs" of research.

Shared mechanisms underlie the control of working memory and attention

Cognitive control guides behaviour by controlling what, when, and how information is represented in the brain¹. For example, attention controls sensory processing; top-down signals from prefrontal and parietal cortex strengthen the representation of task-relevant stimuli^2-4. A similar 'selection' mechanism is thought to control the representations held 'in mind'-in working memory^5-10. Here we show that shared neural mechanisms underlie the selection of items from working memory and attention to sensory stimuli. We trained rhesus monkeys to switch between two tasks, either selecting one item from a set of items held in working memory or attending to one stimulus from a set of visual stimuli. Neural recordings showed that similar representations in prefrontal cortex encoded the control of both selection and attention, suggesting that prefrontal cortex acts as a domain-general controller. By contrast, both attention and selection were represented independently in parietal and visual cortex. Both selection and attention facilitated behaviour by enhancing and transforming the representation of the selected memory or attended stimulus. Specifically, during the selection task, memory items were initially represented in independent subspaces of neural activity in prefrontal cortex. Selecting an item caused its representation to transform from its own subspace to a new subspace used to guide behaviour. A similar transformation occurred for attention. Our results suggest that prefrontal cortex controls cognition by dynamically transforming representations to control what and when cognitive computations are engaged.

The development of retro-cue benefits with extensive practice: Implications for capacity estimation and attentional states in visual working memory

Accessing the contents of visual short-term memory (VSTM) is compromised by information bottlenecks and visual interference between memorization and recall. Retro-cues, displayed after the offset of a memory stimulus and prior to the onset of a probe stimulus, indicate the test item and improve performance in VSTM tasks. It has been proposed that retro-cues aid recall by transferring information from a high-capacity memory store into visual working memory (multiple-store hypothesis). Alternatively, retro-cues could aid recall by redistributing memory resources within the same (low-capacity) working memory store (single-store hypothesis). If retro-cues provide access to a memory store with a capacity exceeding the set size, then, given sufficient training in the use of the retro-cue, near-ceiling performance should be observed. To test this prediction, 10 observers each performed 12 hours across 8 sessions in a retro-cue change-detection task (40,000+ trials total). The results provided clear support for the single-store hypothesis: retro-cue benefits (difference between a condition with and without retro-cues) emerged after a few hundred trials and then remained constant throughout the testing sessions, consistently improving performance by two items, rather than reaching ceiling performance. Surprisingly, we also observed a general increase in performance throughout the experiment in conditions with and without retro-cues, calling into question the generalizability of change-detection tasks in assessing working memory capacity as a stable trait of an observer (data and materials are available at osf.io/9xr82 and github.com/paulzerr/retrocues). In summary, the present findings suggest that retro-cues increase capacity estimates by redistributing memory resources across memoranda within a low-capacity working memory store.

Theoretical distinction between functional states in working memory and their corresponding neural states

Working memory (WM) is important for guiding behaviour, but not always for the next possible action. Here we define a WM item that is currently relevant for guiding behaviour as the functionally "active" item; whereas items maintained in WM, but not immediately relevant to behaviour, are defined as functionally "latent". Traditional neurophysiological theories of WM proposed that content is maintained via persistent neural activity (e.g., stable attractors); however, more recent theories have highlighted the potential role for "activity-silent" mechanisms (e.g., short-term synaptic plasticity). Given these somewhat parallel dichotomies, functionally active and latent cognitive states of WM have been associated with storage based on persistent-activity and activity-silent neural mechanisms, respectively. However, in this article we caution against a one-to-one correspondence between functional and activity states. We argue that the principal theoretical requirement for active and latent WM is that the corresponding neural states play qualitatively different functional roles. We consider a number of candidate solutions, and conclude that the neurophysiological mechanisms for functionally active and latent WM items are theoretically independent of the distinction between persistent activity-based and activity-silent forms of WM storage.

The Dynamic-Processing Model of Working Memory

Recent shifts in the understanding of how the mind and brain retain information in working memory (WM) call for revision to traditional theories. Evidence of dynamic, “activity-silent,” short-term retention processes diverges from conventional models positing that information is always retained in WM by sustained neural activity in buffers. Such evidence comes from machine-learning methods that can decode patterns of brain activity and the simultaneous administration of transcranial magnetic stimulation (TMS) to causally manipulate brain activity in specific areas and time points. TMS can “ping” brain areas to both reactivate latent representations retained in WM and affect memory performance. On the basis of these findings, I argue for a supplement to sustained retention mechanisms. Brain-decoding methods also reveal that dynamic levels of representational codes are retained in WM, and these vary according to task context, from perceptual (sensory) codes in posterior areas to abstract, recoded representations distributed across frontoparietal regions. A dynamic-processing model of WM is advanced to account for the overall pattern of results.

The what, where and how of delay activity

Working memory is characterized by neural activity that persists during the retention interval of delay tasks. Despite the ubiquity of this delay activity across tasks, species and experimental techniques, our understanding of this phenomenon remains incomplete. Although initially there was a narrow focus on sustained activation in a small number of brain regions, methodological and analytical advances have allowed researchers to uncover previously unobserved forms of delay activity various parts of the brain. In light of these new findings, this Review reconsiders what delay activity is, where in the brain it is found, what roles it serves and how it may be generated.

Working memory prioritization impacts neural recovery from distraction

The ability to protect goal-relevant information from disruption over short intervals is a hallmark of working memory. Recent behavioral data suggest that high-priority items in working memory are more vulnerable to disruption. We used functional magnetic resonance imaging to evaluate the hypothesis that prioritization of working memories might impact the recovery of their neural representation(s) after distraction. A delay-period retrospective cue informed participants which of two memory items (a face or a scene) to prioritize during a first delay period. Consistent with prior work, and confirming successful prioritization, multivoxel pattern classifier evidence in perceptual brain regions was higher for cued versus uncued memory items. A distraction task was then imposed before a second retrospective cue informed participants to either "stay" remembering the previously cued item or "switch" to the previously uncued item. This allowed for the evaluation of recovery for high-priority items (on stay trials) and also low-priority items (on switch trials). Classifiers showed successful reinstatement of both high- and low-priority items after distraction, but only low-priority items recovered to their pre-distraction representational levels. Moreover, the degree of prioritization before distraction predicted the amount of disruption for high-priority items after distraction, suggesting that the more a participant prioritized the cued item, the greater the impact of distraction. Our data provide neural evidence that prioritizing working memory information in perceptual regions makes that information more vulnerable to disruption.

Circuit mechanisms for the maintenance and manipulation of information in working memory

Recently it has been proposed that information in working memory (WM) may not always be stored in persistent neuronal activity but can be maintained in 'activity-silent' hidden states, such as synaptic efficacies endowed with short-term synaptic plasticity. To test this idea computationally, we investigated recurrent neural network models trained to perform several WM-dependent tasks, in which WM representation emerges from learning and is not a priori assumed to depend on self-sustained persistent activity. We found that short-term synaptic plasticity can support the short-term maintenance of information, provided that the memory delay period is sufficiently short. However, in tasks that require actively manipulating information, persistent activity naturally emerges from learning, and the amount of persistent activity scales with the degree of manipulation required. These results shed insight into the current debate on WM encoding and suggest that persistent activity can vary markedly between short-term memory tasks with different cognitive demands.

Neural mechanisms of attending to items in working memory

Working memory, the ability to keep recently accessed information available for immediate manipulation, has been proposed to rely on two mechanisms that appear difficult to reconcile: self-sustained neural firing, or the opposite-activity-silent synaptic traces. Here we review and contrast models of these two mechanisms, and then show that both phenomena can co-exist within a unified system in which neurons hold information in both activity and synapses. Rapid plasticity in flexibly-coding neurons allows features to be bound together into objects, with an important emergent property being the focus of attention. One memory item is held by persistent activity in an attended or "focused" state, and is thus remembered better than other items. Other, previously attended items can remain in memory but in the background, encoded in activity-silent synaptic traces. This dual functional architecture provides a unified common mechanism accounting for a diversity of perplexing attention and memory effects that have been hitherto difficult to explain in a single theoretical framework.

Current and future goals are represented in opposite patterns in object-selective cortex

Adaptive behavior requires the separation of current from future goals in working memory. We used fMRI of object-selective cortex to determine the representational (dis)similarities of memory representations serving current and prospective perceptual tasks. Participants remembered an object drawn from three possible categories as the target for one of two consecutive visual search tasks. A cue indicated whether the target object should be looked for first (currently relevant), second (prospectively relevant), or if it could be forgotten (irrelevant). Prior to the first search, representations of current, prospective and irrelevant objects were similar, with strongest decoding for current representations compared to prospective (Experiment 1) and irrelevant (Experiment 2). Remarkably, during the first search, prospective representations could also be decoded, but revealed anti-correlated voxel patterns compared to currently relevant representations of the same category. We propose that the brain separates current from prospective memories within the same neuronal ensembles through opposite representational patterns.

Retrospective Cues Mitigate Information Loss in Human Cortex during Working Memory Storage

Working memory (WM) enables the flexible representation of information over short intervals. It is well established that WM performance can be enhanced by a retrospective cue presented during storage, yet the neural mechanisms responsible for this benefit are unclear. Here, we tested several explanations for retrospective cue benefits by quantifying changes in spatial WM representations reconstructed from alpha-band (8-12 Hz) EEG activity recorded from human participants (both sexes) before and after the presentation of a retrospective cue. This allowed us to track cue-related changes in WM representations with high temporal resolution (tens of milliseconds). Participants encoded the locations of two colored discs for subsequent report. During neutral trials, an uninformative cue instructed participants to remember the locations of both discs across a blank delay, and we observed a monotonic decrease in the fidelity of reconstructed spatial WM representations with time. During valid trials, a 100% reliable cue indicated that the color of the disc participants would be probed to report. Critically, valid cues were presented immediately after the termination of the encoding display ["valid early" (VE) trials] or midway through the delay period ["valid late" (VL) trials]. During VE trials, the gradual loss of location-specific information observed during neutral trials was eliminated, while during VL trials it was partially reversed. Our findings suggest that retrospective cues engage several different mechanisms that together serve to mitigate information loss during WM storage.SIGNIFICANCE STATEMENT Working memory (WM) performance can be improved by a cue presented during storage. This effect, termed a retrospective cue benefit, has been used to explore the limitations of attentional prioritization in WM. However, the mechanisms responsible for retrospective cue benefits are unclear. Here we tested several explanations for retrospective cue benefits by examining how they influence WM representations reconstructed from human EEG activity. This approach allowed us to visualize, quantify, and track the effects of retrospective cues with high temporal resolution (on the order of tens of milliseconds). We show that under different circumstances retrospective cues can both eliminate and even partially reverse information loss during WM storage, suggesting that retrospective cue benefits have manifold origins.

The removal of information from working memory

What happens to goal-relevant information in working memory after it is no longer needed? Here, we review evidence for a selective removal process that operates on outdated information to limit working memory load and hence facilitates the maintenance of goal-relevant information. Removal alters the representations of irrelevant content so as to reduce access to it, thereby improving access to the remaining relevant content and also facilitating the encoding of new information. Both behavioral and neural evidence support the existence of a removal process that is separate from forgetting due to decay or interference. We discuss the potential mechanisms involved in removal and characterize the time course and duration of the process. In doing so, we propose the existence of two forms of removal: one is temporary, and reversible, which modifies working memory content without impacting content-to-context bindings, and another is permanent, which unbinds the content from its context in working memory (without necessarily impacting long-term forgetting). Finally, we discuss limitations on removal and prescribe conditions for evaluating evidence for or against this process.

Drift in Neural Population Activity Causes Working Memory to Deteriorate Over Time

Short-term memories are thought to be maintained in the form of sustained spiking activity in neural populations. Decreases in recall precision observed with increasing number of memorized items can be accounted for by a limit on total spiking activity, resulting in fewer spikes contributing to the representation of each individual item. Longer retention intervals likewise reduce recall precision, but it is unknown what changes in population activity produce this effect. One possibility is that spiking activity becomes attenuated over time, such that the same mechanism accounts for both effects of set size and retention duration. Alternatively, reduced performance may be caused by drift in the encoded value over time, without a decrease in overall spiking activity. Human participants of either sex performed a variable-delay cued recall task with a saccadic response, providing a precise measure of recall latency. Based on a spike integration model of decision making, if the effects of set size and retention duration are both caused by decreased spiking activity, we would predict a fixed relationship between recall precision and response latency across conditions. In contrast, the drift hypothesis predicts no systematic changes in latency with increasing delays. Our results show both an increase in latency with set size, and a decrease in response precision with longer delays within each set size, but no systematic increase in latency for increasing delay durations. These results were quantitatively reproduced by a model based on a limited neural resource in which working memories drift rather than decay with time.

SIGNIFICANCE STATEMENT Rapid deterioration over seconds is a defining feature of short-term memory, but what mechanism drives this degradation of internal representations? Here, we extend a successful population coding model of working memory by introducing possible mechanisms of delay effects. We show that a decay in neural signal over time predicts that the time required for memory retrieval will increase with delay, whereas a random drift in the stored value predicts no effect of delay on retrieval time. Testing these predictions in a multi-item memory task with an eye movement response, we identified drift as a key mechanism of memory decline. These results provide evidence for a dynamic spiking basis for working memory, in contrast to recent proposals of activity-silent storage.

Behavioral decoding of working memory items inside and outside the focus of attention

How we attend to our thoughts affects how we attend to our environment. Holding information in working memory can automatically bias visual attention toward matching information. By observing attentional biases on reaction times to visual search during a memory delay, it is possible to reconstruct the source of that bias using machine learning techniques and thereby behaviorally decode the content of working memory. Can this be done when more than one item is held in working memory? There is some evidence that multiple items can simultaneously bias attention, but the effects have been inconsistent. One explanation may be that items are stored in different states depending on the current task demands. Recent models propose functionally distinct states of representation for items inside versus outside the focus of attention. Here, we use behavioral decoding to evaluate whether multiple memory items-including temporarily irrelevant items outside the focus of attention-exert biases on visual attention. Only the single item in the focus of attention was decodable. The other item showed a brief attentional bias that dissipated until it returned to the focus of attention. These results support the idea of dynamic, flexible states of working memory across time and priority.

Cortical specialization for attended versus unattended working memory

Items held in working memory can be either attended or not, depending on their current behavioral relevance. It has been suggested that unattended contents might be solely retained in an activity-silent form. Instead, we demonstrate here that encoding unattended contents involves a division of labor. While visual cortex only maintains attended items, intraparietal areas and the frontal eye fields represent both attended and unattended items.