Stochastic sampling provides a unifying account of visual working memory limits.

Research into human working memory limits has been shaped by the competition between different formal models, with a central point of contention being whether internal representations are continuous or discrete. Here we describe a sampling approach derived from principles of neural coding as a framework to understand working memory limits. Reconceptualizing existing models in these terms reveals strong commonalities between seemingly opposing accounts, but also allows us to identify specific points of difference. We show that the discrete versus continuous nature of sampling is not critical to model fits, but that, instead, random variability in sample counts is the key to reproducing human performance in both single- and whole-report tasks. A probabilistic limit on the number of items successfully retrieved is an emergent property of stochastic sampling, requiring no explicit mechanism to enforce it. These findings resolve discrepancies between previous accounts and establish a unified computational framework for working memory that is compatible with neural principles.

Swap errors in visual working memory are fully explained by cue-feature variability.

In cue-based recall from working memory, incorrectly reporting features of an uncued item may be referred to as a "swap" error. One account of these errors ascribes them to variability in memory for the cue features leading to erroneous selection of a non-target item, especially if it is similar to the target in the cue-feature dimension. However, alternative accounts of swap errors include cue-independent misbinding, and strategic guessing when the cued item is not in memory. Here we investigated the cause of swap errors by manipulating the variability with which either cue or report features (orientations in Exp 1; motion directions in Exp 2) were encoded. We found that swap errors increased with increasing variability in memory for the cue features, and their changing frequency could be quantitatively predicted based on recall variability when the same feature was used for report. These results are inconsistent with the hypothesis that swaps are a strategic response to forgotten items, and suggest that swap errors could be wholly accounted for by confusions due to cue-dimension variability. In a third experiment we examined whether spatial configuration of memory arrays in tasks with spatial cueing has an influence on swap error frequency. We observed a specific tendency to make swap errors to non-targets located precisely opposite to the cued location, suggesting that stimulus positions are partially encoded in a non-metric format.

Consequence of stroke for feature recall and binding in visual working memory.

Visual memory for objects involves the integration, or binding, of individual features into a coherent representation. We used a novel approach to assess feature binding, using a delayed-reproduction task in combination with computational modeling and lesion analysis. We assessed stroke patients and neurotypical controls on a visual working memory task in which spatial arrays of colored disks were presented. After a brief delay, participants either had to report the color of one disk cued by its location or the location of one disk cued by its color. Our results demonstrate that, in the controls, report imprecision and swap errors (non-target reports) can be explained by a single source of variability. Stroke patients showed an overall decrease in memory precision for both color and location, with only limited evidence for deviations from the predicted relationship between report precision and swap errors. These deviations were primarily deficits in reporting items rather than selecting items based on the cue. Atlas-based lesion-symptom mapping showed that selection and reporting deficits, precision in reporting color, and precision in reporting location were associated with different lesion profiles. Deficits in binding are associated with lesions in the left somatosensory cortex, deficits in the precision of reporting color with bilateral fronto-parietal regions, and no anatomical substrates were identified for precision in reporting location. Our results converge with previous reports that working memory representations are widely distributed in the brain and can be found across sensory, parietal, temporal, and prefrontal cortices. Stroke patients demonstrate mostly subtle impairments in visual working memory, perhaps because representations from different areas in the brain can partly compensate for impaired encoding in lesioned areas. These findings contribute to understanding of the relation between memorizing features and their bound representations.

Mechanisms of feature binding in visual working memory are stable over long delays.

The ability to accurately retain the binding between the features of different objects is a critical element of visual working memory. The underlying mechanism can be elucidated by analyzing correlations of response errors in dual-report experiments, in which participants have to report two features of a single item from a previously viewed stimulus array. Results from separate previous studies using different cueing conditions have indicated that location takes a privileged role in mediating binding between other features, in that largely independent response errors have been observed when location was used as a cue, but errors were highly correlated when location was one of the reported features. Earlier results from change detection tasks likewise support such a special role of location, but they also suggest that this role is substantially reduced for longer retention intervals in favor of object-based representation. In the present study, we replicated the findings of previous dual-report tasks with different cueing conditions, using matched stimuli and procedures. Moreover, we show that the observed patterns of error correlations remain qualitatively unchanged with longer retention intervals. Fits with neural population models demonstrate that the behavioral results at long, as well as short, delays are best explained by memory representations in independent feature maps, in which an item's features are bound to each other only via their shared location.

Transsaccadic integration operates independently in different feature dimensions.

Our knowledge about objects in our environment reflects an integration of current visual input with information from preceding gaze fixations. Such a mechanism may reduce uncertainty but requires the visual system to determine which information obtained in different fixations should be combined or kept separate. To investigate the basis of this decision, we conducted three experiments. Participants viewed a stimulus in their peripheral vision and then made a saccade that shifted the object into the opposite hemifield. During the saccade, the object underwent changes of varying magnitude in two feature dimensions (Experiment 1, color and location; Experiments 2 and 3, color and orientation). Participants reported whether they detected any change and estimated one of the postsaccadic features. Integration of presaccadic with postsaccadic input was observed as a bias in estimates toward the presaccadic feature value. In all experiments, presaccadic bias weakened as the magnitude of the transsaccadic change in the estimated feature increased. Changes in the other feature, despite having a similar probability of detection, had no effect on integration. Results were quantitatively captured by an observer model where the decision whether to integrate information from sequential fixations is made independently for each feature and coupled to awareness of a feature change.

Transsaccadic integration relies on a limited memory resource.

Saccadic eye movements cause large-scale transformations of the image falling on the retina. Rather than starting visual processing anew after each saccade, the visual system combines post-saccadic information with visual input from before the saccade. Crucially, the relative contribution of each source of information is weighted according to its precision, consistent with principles of optimal integration. We reasoned that, if pre-saccadic input is maintained in a resource-limited store, such as visual working memory, its precision will depend on the number of items stored, as well as their attentional priority. Observers estimated the color of stimuli that changed imperceptibly during a saccade, and we examined where reports fell on the continuum between pre- and post-saccadic values. Bias toward the post-saccadic color increased with the set size of the pre-saccadic display, consistent with an increased weighting of the post-saccadic input as precision of the pre-saccadic representation declined. In a second experiment, we investigated if transsaccadic memory resources are preferentially allocated to attentionally prioritized items. An arrow cue indicated one pre-saccadic item as more likely to be chosen for report. As predicted, valid cues increased response precision and biased responses toward the pre-saccadic color. We conclude that transsaccadic integration relies on a limited memory resource that is flexibly distributed between pre-saccadic stimuli.

Independent working memory resources for egocentric and allocentric spatial information.

Visuospatial working memory enables us to maintain access to visual information for processing even when a stimulus is no longer present, due to occlusion, our own movements, or transience of the stimulus. Here we show that, when localizing remembered stimuli, the precision of spatial recall does not rely solely on memory for individual stimuli, but additionally depends on the relative distances between stimuli and visual landmarks in the surroundings. Across three separate experiments, we consistently observed a spatially selective improvement in the precision of recall for items located near a persistent landmark. While the results did not require that the landmark be visible throughout the memory delay period, it was essential that it was visible both during encoding and response. We present a simple model that can accurately capture human performance by considering relative (allocentric) spatial information as an independent localization estimate which degrades with distance and is optimally integrated with egocentric spatial information. Critically, allocentric information was encoded without cost to egocentric estimation, demonstrating independent storage of the two sources of information. Finally, when egocentric and allocentric estimates were put in conflict, the model successfully predicted the resulting localization errors. We suggest that the relative distance between stimuli represents an additional, independent spatial cue for memory recall. This cue information is likely to be critical for spatial localization in natural settings which contain an abundance of visual landmarks.

Efficient Coding in Visual Working Memory Accounts for Stimulus-Specific Variations in Recall.

Recall of visual features from working memory varies in both bias and precision depending on stimulus parameters. Whereas a number of models can approximate the average distribution of recall error across target stimuli, attempts to model how error varies with the choice of target have been ad hoc Here we adapt a neural model of working memory to provide a principled account of these stimulus-specific effects, by allowing each neuron's tuning function to vary according to the principle of efficient coding, which states that neural responses should be optimized with respect to the frequency of stimuli in nature. For orientation, this means incorporating a prior that favors cardinal over oblique orientations. While continuing to capture the changes in error distribution with set size, the resulting model accurately described stimulus-specific variations as well, better than a slot-based competitor. Efficient coding produces a repulsive bias away from cardinal orientations, a bias that ought to be sensitive to changes in the environmental statistics. We subsequently tested whether shifts in the stimulus distribution influenced response bias to uniformly sampled target orientations in human subjects (of either sex). Across adaptation blocks, we manipulated the distribution of nontarget items by sampling from a bimodal congruent (incongruent) distribution with peaks centered on cardinal (oblique) orientations. Preadaptation responses were repulsed away from the cardinal axes. However, exposure to the incongruent distribution produced systematic decreases in repulsion that persisted after adaptation. This result confirms the role of prior expectation in generating stimulus-specific effects and validates the neural framework.SIGNIFICANCE STATEMENT Theories of neural coding have been used successfully to explain how errors in recall from working memory depend on the number of items stored. However, recall of visual features also shows stimulus-specific variation in bias and precision. Here we unify two previously unconnected theories, the neural resource model of working memory and the efficient coding framework, to provide a principled account of these stimulus-specific effects. Given the importance of working memory limitations to multiple aspects of human and animal behavior, and the recent high-profile advances in theories of efficient coding, our modeling framework provides a richer, yet parsimonious, description of how orientation encoding influences visual working memory performance.

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.

Visual Working Memory Is Independent of the Cortical Spacing Between Memoranda.

The sensory recruitment hypothesis states that visual short-term memory is maintained in the same visual cortical areas that initially encode a stimulus' features. Although it is well established that the distance between features in visual cortex determines their visibility, a limitation known as crowding, it is unknown whether short-term memory is similarly constrained by the cortical spacing of memory items. Here, we investigated whether the cortical spacing between sequentially presented memoranda affects the fidelity of memory in humans (of both sexes). In a first experiment, we varied cortical spacing by taking advantage of the log-scaling of visual cortex with eccentricity, presenting memoranda in peripheral vision sequentially along either the radial or tangential visual axis with respect to the fovea. In a second experiment, we presented memoranda sequentially either within or beyond the critical spacing of visual crowding, a distance within which visual features cannot be perceptually distinguished due to their nearby cortical representations. In both experiments and across multiple measures, we found strong evidence that the ability to maintain visual features in memory is unaffected by cortical spacing. These results indicate that the neural architecture underpinning working memory has properties inconsistent with the known behavior of sensory neurons in visual cortex. Instead, the dissociation between perceptual and memory representations supports a role of higher cortical areas such as posterior parietal or prefrontal regions or may involve an as yet unspecified mechanism in visual cortex in which stimulus features are bound to their temporal order.SIGNIFICANCE STATEMENT Although much is known about the resolution with which we can remember visual objects, the cortical representation of items held in short-term memory remains contentious. A popular hypothesis suggests that memory of visual features is maintained via the recruitment of the same neural architecture in sensory cortex that encodes stimuli. We investigated this claim by manipulating the spacing in visual cortex between sequentially presented memoranda such that some items shared cortical representations more than others while preventing perceptual interference between stimuli. We found clear evidence that short-term memory is independent of the intracortical spacing of memoranda, revealing a dissociation between perceptual and memory representations. Our data indicate that working memory relies on different neural mechanisms from sensory perception.

Internal but not external noise frees working memory resources.

The precision with which visual information can be recalled from working memory declines as the number of items in memory increases. This finding has been explained in terms of the distribution of a limited representational resource between items. Here we investigated how the sensory strength of memoranda affects resource allocation. We manipulated signal strength of an orientation stimulus in two ways: we varied the internal (sensory) noise by adjusting stimulus contrast, and varied the external (stimulus) noise by altering the within-stimulus variability. Both manipulations had similar effects on the precision with which the orientation could be recalled, but differed in their impact on memory for other stimuli. These results indicate that increasing internal noise released resources that could be used to store other stimuli more precisely; increasing external noise had no such effect. We show that these observations can be captured by a simple neural model of working memory encoding, in which spiking activity takes on the role of the limited resource.

Reassessing the Evidence for Capacity Limits in Neural Signals Related to Working Memory.

In 2004, two landmark studies described the discovery of brain imaging (functional magnetic resonance imaging and electroencephalography) signals that increase with the number of items held in visual working memory (WM). These studies claimed that the signals leveled off (plateaued) once the number of memoranda reached the capacity of WM, as estimated by the prevailing model of the time. However, alternative models were not considered, and changing concepts of WM in the more than a decade since these studies were published necessitate a re-evaluation of their findings; newer models that provide the most accurate account of behavioral data do not incorporate a fixed limit on the number of items stored. Furthermore, an important claim made about the original studies, that signals plateau at each individual's estimated capacity, has never been tested. Here, we pit the plateau model of signal strength against an alternative, saturation model, a biophysically plausible account in which signals increase continuously without plateau. We show that the saturation model provides a better description of the original data, challenging the assumption that imaging results provide evidence for a fixed item limit in WM.

Recall of facial expressions and simple orientations reveals competition for resources at multiple levels of the visual hierarchy.

Many studies of visual working memory have tested humans' ability to reproduce primary visual features of simple objects, such as the orientation of a grating or the hue of a color patch, following a delay. A consistent finding of such studies is that precision of responses declines as the number of items in memory increases. Here we compared visual working memory for primary features and high-level objects. We presented participants with memory arrays consisting of oriented gratings, facial expressions, or a mixture of both. Precision of reproduction for all facial expressions declined steadily as the memory load was increased from one to five faces. For primary features, this decline and the specific distributions of error observed, have been parsimoniously explained in terms of neural population codes. We adapted the population coding model for circular variables to the non-circular and bounded parameter space used for expression estimation. Total population activity was held constant according to the principle of normalization and the intensity of expression was decoded by drawing samples from the Bayesian posterior distribution. The model fit the data well, showing that principles of population coding can be applied to model memory representations at multiple levels of the visual hierarchy. When both gratings and faces had to be remembered, an asymmetry was observed. Increasing the number of faces decreased precision of orientation recall, but increasing the number of gratings did not affect recall of expression, suggesting that memorizing faces involves the automatic encoding of low-level features, in addition to higher-level expression information.

Neural Architecture for Feature Binding in Visual Working Memory.

Binding refers to the operation that groups different features together into objects. We propose a neural architecture for feature binding in visual working memory that employs populations of neurons with conjunction responses. We tested this model using cued recall tasks, in which subjects had to memorize object arrays composed of simple visual features (color, orientation, and location). After a brief delay, one feature of one item was given as a cue, and the observer had to report, on a continuous scale, one or two other features of the cued item. Binding failure in this task is associated with swap errors, in which observers report an item other than the one indicated by the cue. We observed that the probability of swapping two items strongly correlated with the items' similarity in the cue feature dimension, and found a strong correlation between swap errors occurring in spatial and nonspatial report. The neural model explains both swap errors and response variability as results of decoding noisy neural activity, and can account for the behavioral results in quantitative detail. We then used the model to compare alternative mechanisms for binding nonspatial features. We found the behavioral results fully consistent with a model in which nonspatial features are bound exclusively via their shared location, with no indication of direct binding between color and orientation. These results provide evidence for a special role of location in feature binding, and the model explains how this special role could be realized in the neural system.SIGNIFICANCE STATEMENT The problem of feature binding is of central importance in understanding the mechanisms of working memory. How do we remember not only that we saw a red and a round object, but that these features belong together to a single object rather than to different objects in our environment? Here we present evidence for a neural mechanism for feature binding in working memory, based on encoding of visual information by neurons that respond to the conjunction of features. We find clear evidence that nonspatial features are bound via space: we memorize directly where a color or an orientation appeared, but we memorize which color belonged with which orientation only indirectly by virtue of their shared location.

Fidelity of the representation of value in decision-making.

The ability to make optimal decisions depends on evaluating the expected rewards associated with different potential actions. This process is critically dependent on the fidelity with which reward value information can be maintained in the nervous system. Here we directly probe the fidelity of value representation following a standard reinforcement learning task. The results demonstrate a previously-unrecognized bias in the representation of value: extreme reward values, both low and high, are stored significantly more accurately and precisely than intermediate rewards. The symmetry between low and high rewards pertained despite substantially higher frequency of exposure to high rewards, resulting from preferential exploitation of more rewarding options. The observed variation in fidelity of value representation retrospectively predicted performance on the reinforcement learning task, demonstrating that the bias in representation has an impact on decision-making. A second experiment in which one or other extreme-valued option was omitted from the learning sequence showed that representational fidelity is primarily determined by the relative position of an encoded value on the scale of rewards experienced during learning. Both variability and guessing decreased with the reduction in the number of options, consistent with allocation of a limited representational resource. These findings have implications for existing models of reward-based learning, which typically assume defectless representation of reward value.

Failure of self-consistency in the discrete resource model of visual working memory.

The discrete resource model of working memory proposes that each individual has a fixed upper limit on the number of items they can store at one time, due to division of memory into a few independent "slots". According to this model, responses on short-term memory tasks consist of a mixture of noisy recall (when the tested item is in memory) and random guessing (when the item is not in memory). This provides two opportunities to estimate capacity for each observer: first, based on their frequency of random guesses, and second, based on the set size at which the variability of stored items reaches a plateau. The discrete resource model makes the simple prediction that these two estimates will coincide. Data from eight published visual working memory experiments provide strong evidence against such a correspondence. These results present a challenge for discrete models of working memory that impose a fixed capacity limit.

A neural model of retrospective attention in visual working memory.

An informative cue that directs attention to one of several items in working memory improves subsequent recall of that item. Here we examine the mechanism of this retro-cue effect using a model of short-term memory based on neural population coding. Our model describes recalled feature values as the output of an optimal decoding of spikes generated by a tuned population of neurons. This neural model provides a better account of human recall data than an influential model that assumes errors can be described as a mixture of normally distributed noise and random guesses. The retro-cue benefit is revealed to be consistent with a higher firing rate of the population encoding the cued versus uncued items, with no difference in tuning specificity. Additionally, a retro-cued item is less likely to be swapped with another item in memory, an effect that can also be explained by greater activity of the underlying population. These results provide a parsimonious account of the effects of retrospective attention on recall and demonstrate a principled method for investigating neural representations with behavioral tasks.

Reduced Hippocampal Functional Connectivity During Episodic Memory Retrieval in Autism.

Increasing recent research has sought to understand the recollection impairments experienced by individuals with autism spectrum disorder (ASD). Here, we tested whether these memory deficits reflect a reduction in the probability of retrieval success or in the precision of memory representations. We also used functional magnetic resonance imaging (fMRI) to study the neural mechanisms underlying memory encoding and retrieval in ASD, focusing particularly on the functional connectivity of core episodic memory networks. Adults with ASD and typical control participants completed a memory task that involved studying visual displays and subsequently using a continuous dial to recreate their appearance. The ASD group exhibited reduced retrieval success, but there was no evidence of a difference in retrieval precision. fMRI data revealed similar patterns of brain activity and functional connectivity during memory encoding in the 2 groups, though encoding-related lateral frontal activity predicted subsequent retrieval success only in the control group. During memory retrieval, the ASD group exhibited attenuated lateral frontal activity and substantially reduced hippocampal connectivity, particularly between hippocampus and regions of the fronto-parietal control network. These findings demonstrate notable differences in brain function during episodic memory retrieval in ASD and highlight the importance of functional connectivity to understanding recollection-related retrieval deficits in this population.

Restoration of fMRI Decodability Does Not Imply Latent Working Memory States.

Recent imaging studies have challenged the prevailing view that working memory is mediated by sustained neural activity. Using machine learning methods to reconstruct memory content, these studies found that previously diminished representations can be restored by retrospective cueing or other forms of stimulation. These findings have been interpreted as evidence for an activity-silent working memory state that can be reactivated dependent on task demands. Here, we test the validity of this conclusion by formulating a neural process model of working memory based on sustained activity and using this model to emulate a spatial recall task with retro-cueing. The simulation reproduces both behavioral and fMRI results previously taken as evidence for latent states, in particular the restoration of spatial reconstruction quality following an informative cue. Our results demonstrate that recovery of the decodability of an imaging signal does not provide compelling evidence for an activity-silent working memory state.

Automatic and intentional influences on saccade landing.

Saccadic eye movements enable us to rapidly direct our high-resolution fovea onto relevant parts of the visual world. However, while we can intentionally select a location as a saccade target, the wider visual scene also influences our executed movements. In the presence of multiple objects, eye movements may be "captured" to the location of a distractor object, or be biased toward the intermediate position between objects (the "global effect"). Here we examined how the relative strengths of the global effect and visual object capture changed with saccade latency, the separation between visual items and stimulus contrast. Importantly, while many previous studies have omitted giving observers explicit instructions, we instructed participants to either saccade to a specified target object or to the midpoint between two stimuli. This allowed us to examine how their explicit movement goal influenced the likelihood that their saccades terminated at either the target, distractor, or intermediate locations. Using a probabilistic mixture model, we found evidence that both visual object capture and the global effect co-occurred at short latencies and declined as latency increased. As object separation increased, capture came to dominate the landing positions of fast saccades, with reduced global effect. Using the mixture model fits, we dissociated the proportion of unavoidably captured saccades to each location from those intentionally directed to the task goal. From this we could extract the time course of competition between automatic capture and intentional targeting. We show that task instructions substantially altered the distribution of saccade landing points, even at the shortest latencies.NEW & NOTEWORTHY When making an eye movement to a target location, the presence of a nearby distractor can cause the saccade to unintentionally terminate at the distractor itself or the average position in between stimuli. With probabilistic mixture models, we quantified how both unavoidable capture and goal-directed targeting were influenced by changing the task and the target-distractor separation. Using this novel technique, we could extract the time course over which automatic and intentional processes compete for control of saccades.