Introspective judgments predict the precision and likelihood of successful maintenance of visual working memory

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.

Neural mechanisms of resource allocation in working memory

To mitigate capacity limits of working memory, people allocate resources according to an item's relevance. However, the neural mechanisms supporting such a critical operation remain unknown. Here, we developed computational neuroimaging methods to decode and demix neural responses associated with multiple items in working memory with different priorities. In striate and extrastriate cortex, the gain of neural responses tracked the priority of memoranda. Higher-priority memoranda were decoded with smaller error and lower uncertainty. Moreover, these neural differences predicted behavioral differences in memory prioritization. Remarkably, trialwise variability in the magnitude of delay activity in frontal cortex predicted differences in decoded precision between low and high-priority items in visual cortex. These results suggest a model in which feedback signals broadcast from frontal cortex sculpt the gain of memory representations in visual cortex according to behavioral relevance, thus, identifying a neural mechanism for resource allocation.

I remember it now, so I'll remember it later: Working memory strength guides predictions for long-term memory performance

Judgments of learning (JOLs) are assumed to be made inferentially, based on cues. This cue-utilization approach substituted the theory that memory strength guides JOLs. The rejection of this theory ignores the existence of two memory systems: working memory (WM), which holds representations immediately accessible, and long-term memory (LTM), which is a permanent store. By manipulating and measuring WM strength, we tested a revised version of the memory-strength theory in which JOLs are guided by WM representations. In Experiment 1, participants memorized sequences of two or four colored objects, then they provided JOLs for an LTM test of these objects, and performed a WM test on the objects' colors. After learning 200 objects, the LTM test followed. Sequence-length affected WM, but not LTM performance. JOLs, however, were higher for sequences of two than for four objects and correlated higher with WM than LTM performance. We replicated these results with a simultaneous presentation of the objects (Experiment 2), in the absence of a WM test (Experiment 3), and in a word-pair task (Experiment 4). Overall, our findings are consistent with the revised memory-strength theory. WM strength should therefore be considered when examining the factors guiding JOLs.

Probabilistic and rich individual working memories revealed by a betting game

When asked to remember a color, do people remember a point estimate (e.g., a particular shade of red), a point estimate plus an uncertainty estimate, or are memory representations rich probabilistic distributions over feature space? We asked participants to report the color of a circle held in working memory. Rather than collecting a single report per trial, we had participants place multiple bets to create trialwise uncertainty distributions. Bet dispersion correlated with performance, indicating that internal uncertainty guided bet placement. While the first bet was on average the most precisely placed, the later bets systematically shifted the distribution closer to the target, resulting in asymmetrical distributions about the first bet. This resulted in memory performance improvements when averaging across bets, and overall suggests that memory representations contain more information than can be conveyed by a single response. The later bets contained target information even when the first response would generally be classified as a guess or report of an incorrect item, suggesting that such failures are not all-or-none. This paradigm provides multiple pieces of evidence that memory representations are rich and probabilistic. Crucially, standard discrete response paradigms underestimate the amount of information in memory representations.

Studying the neural representations of uncertainty

The study of the brain's representations of uncertainty is a central topic in neuroscience. Unlike most quantities of which the neural representation is studied, uncertainty is a property of an observer's beliefs about the world, which poses specific methodological challenges. We analyze how the literature on the neural representations of uncertainty addresses those challenges and distinguish between 'code-driven' and 'correlational' approaches. Code-driven approaches make assumptions about the neural code for representing world states and the associated uncertainty. By contrast, correlational approaches search for relationships between uncertainty and neural activity without constraints on the neural representation of the world state that this uncertainty accompanies. To compare these two approaches, we apply several criteria for neural representations: sensitivity, specificity, invariance and functionality. Our analysis reveals that the two approaches lead to different but complementary findings, shaping new research questions and guiding future experiments.

A spatially continuous diffusion model of visual working memory

We present results from five visual working memory (VWM) experiments in which participants were briefly shown between 2 and 6 colored squares. They were then cued to recall the color of one of the squares and they responded by choosing the color on a continuous color wheel. The experiments provided response proportions and response time (RT) measures as a function of angle for the choices. Current VWM models for this task include discrete models that assume an item is either within working memory or not and resource models that assume that memory strength varies as a function of the number of items. Because these models do not include processes that allow them to account for RT data, we implemented them within the spatially continuous diffusion model (SCDM, Ratcliff, 2018) and use the experimental data to evaluate these combined models. In the SCDM, evidence retrieved from memory is represented as a spatially continuous normal distribution and this drives the decision process until a criterion (represented as a 1-D line) is reached, which produces a decision. Noise in the accumulation process is represented by continuous Gaussian process noise over spatial position. The models that fit best from the discrete and resource-based classes converged on a common model that had a guessing component and that allowed the height of the normal memory-strength distribution to vary with number of items. The guessing component was implemented as a regular decision process driven by a flat evidence distribution, a zero-drift process. The combination of choice and RT data allows models that were not identifiable based on choice data alone to be discriminated.

Modality independent or modality specific? Common computations underlie confidence judgements in visual and auditory decisions

The mechanisms that enable humans to evaluate their confidence across a range of different decisions remain poorly understood. To bridge this gap in understanding, we used computational modelling to investigate the processes that underlie confidence judgements for perceptual decisions and the extent to which these computations are the same in the visual and auditory modalities. Participants completed two versions of a categorisation task with visual or auditory stimuli and made confidence judgements about their category decisions. In each modality, we varied both evidence strength, (i.e., the strength of the evidence for a particular category) and sensory uncertainty (i.e., the intensity of the sensory signal). We evaluated several classes of computational models which formalise the mapping of evidence strength and sensory uncertainty to confidence in different ways: 1) unscaled evidence strength models, 2) scaled evidence strength models, and 3) Bayesian models. Our model comparison results showed that across tasks and modalities, participants take evidence strength and sensory uncertainty into account in a way that is consistent with the scaled evidence strength class. Notably, the Bayesian class provided a relatively poor account of the data across modalities, particularly in the more complex categorisation task. Our findings suggest that a common process is used for evaluating confidence in perceptual decisions across domains, but that the parameter settings governing the process are tuned differently in each modality. Overall, our results highlight the impact of sensory uncertainty on confidence and the unity of metacognitive processing across sensory modalities.

Metacognition in working memory: Confidence judgments during an n-back task

Metacognition in working memory (WM) has received less attention than episodic memory, and few studies have investigated confidence judgements while carrying out a verbal WM task. The present study investigated whether individuals are aware of their own level of performance while carrying out an ongoing verbal WM task, and whether judgments of confidence are sensitive to factors that determine WM performance. A verbal n-back task was adapted to obtain confidence judgments on a trial-by-trial basis. Memory load and lure interference were manipulated. Results showed that metacognition judgments were affected by memory load and levels of interference just as performance accuracy. Even when judgments were sensitive to memory factors, participants were overconfident and generally showed poor metacognitive accuracy at discriminating between erroneous and accurate responses. Results are discussed in terms of possible cues contributing to metacognitive judgements during an ongoing WM task and reasons for WM metacognitive accuracy.

Bayesian inference in ring attractor networks

Working memories are thought to be held in attractor networks in the brain. These attractors should keep track of the uncertainty associated with each memory, so as to weigh it properly against conflicting new evidence. However, conventional attractors do not represent uncertainty. Here, we show how uncertainty could be incorporated into an attractor, specifically a ring attractor that encodes head direction. First, we introduce a rigorous normative framework (the circular Kalman filter) for benchmarking the performance of a ring attractor under conditions of uncertainty. Next, we show that the recurrent connections within a conventional ring attractor can be retuned to match this benchmark. This allows the amplitude of network activity to grow in response to confirmatory evidence, while shrinking in response to poor-quality or strongly conflicting evidence. This "Bayesian ring attractor" performs near-optimal angular path integration and evidence accumulation. Indeed, we show that a Bayesian ring attractor is consistently more accurate than a conventional ring attractor. Moreover, near-optimal performance can be achieved without exact tuning of the network connections. Finally, we use large-scale connectome data to show that the network can achieve near-optimal performance even after we incorporate biological constraints. Our work demonstrates how attractors can implement a dynamic Bayesian inference algorithm in a biologically plausible manner, and it makes testable predictions with direct relevance to the head direction system as well as any neural system that tracks direction, orientation, or periodic rhythms.

Awareness of the relative quality of spatial working memory representations

Working memory (WM) is the ability to maintain and manipulate information no longer accessible in the environment. The brain maintains WM representations over delay periods in noisy population-level activation patterns, resulting in variability in WM representations across items and trials. It is established that participants can introspect aspects of the quality of WM representations, and that they can accurately compare which of several WM representations of stimulus features like orientation or color is better on each trial. However, whether this ability to evaluate and compare the quality of multiple WM representations extends to spatial WM tasks remains unknown. Here, we employed a memory-guided saccade task to test recall errors for remembered spatial locations when participants were allowed to choose the most precise representation to report. Participants remembered either one or two spatial locations over a delay and reported one item's location with a saccade. On trials with two spatial locations, participants reported either the spatial location of a randomly cued item, or the location of the stimulus they remembered best. We found a significant improvement in recall error and increase in response time (RT) when participants reported their best-remembered item compared with trials in which they were randomly cued. These results demonstrate that participants can accurately introspect the relative quality of neural WM representations for spatial position, consistent with previous observations for other stimulus features, and support a model of WM coding involving noisy representations across items and trials.

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.

Little Support for Discrete Item Limits in Visual Working Memory

Some theorists argue that working memory is limited to a discrete number of items and that additional items are not encoded at all. Adam et al. (2017) presented evidence supporting this hypothesis: Participants reproduced visual features of up to six items in a self-chosen order. After the third or fourth response, error distributions were indistinguishable from guessing. I present four experiments with young adults (each N = 24) reexamining this finding. Experiment 1 presented items slowly and sequentially. Experiment 2 presented them simultaneously but longer than in the experiments of Adam et al. Experiments 3 and 4 exactly replicated one original experiment of Adam et al. All four experiments failed to replicate the evidence for guessing-like error distributions. Modeling data from individuals revealed a mixture of some who do and others who do not produce guessing-like distributions. This heterogeneity increases the credibility of an alternative to the item-limit hypothesis: Some individuals decide to guess on hard trials even when they have weak information in memory.

Trial-by-trial mouse trajectory predicts variance in precision across working memory representations: A critical reanalysis of Hao et al. (2021)

Multiple representations in visual working memory (VWM) can vary in mnemonic precision. This inhomogeneity of VWM precision has received some support from recent studies with the whole-report procedure, in which all memory items are recalled in free or forced orders. Recently, Hao et al. (2021, Cognition, 214, 104739) added a novel item-selection stage before each memory recall and found smaller between-trial variance in mouse trajectory during the selection stage in free-recall condition as compared with forced recall, which was taken as evidence for less between-item interference and the resulting precision benefit under free recall. Here, we reanalyzed the original dataset with a different analytic approach and attempted independent hypothesis testing focusing on within-trial trajectory deviations. We found that the direction of trial-by-trial trajectory bias for the first to-be-recalled item was predictive of the relative mnemonic precision of the remaining items. Critically, this relationship was only present for forced recall but not for free recall. Hierarchical Bayesian modeling of recall errors further identified that this relationship was selectively driven by VWM precision. Together, our reanalysis provides evidence for the source of between-item interference and its direct association with variable precision of VWM representations, and further highlights the novel methodological benefits of probing memory decisional processes using mouse trajectory data.

Tracking attentional states: Assessing the relationship between sustained and selective focused attention in visual working memory

Attention has multiple influences on visual working memory (VWM). Fluctuations in sustained attention predict VWM performance. Furthermore, focusing selective attention in VWM by retro-cuing the to-be-tested item during maintenance boosts retrieval. So far, we lack knowledge how the ability to focus selective attention relates to the state of sustained attention during the VWM trial. Here, we combined a retro-cue task and a self-rated attention protocol to test whether focusing selective attention via retro-cues: (1) mitigates spontaneous attention fluctuations, in which case retro-cues should be more helpful under low levels of self-rated attention; (2) depends on an optimal state of sustained attention, in which case retro-cue benefits should be largest under high levels of self-rated attention; or (3) is independent of sustained attention, in which case retro-cue benefits and self-rated attention effects should be additive. Our data supported the additive hypothesis. Across four experiments, self-rated attention levels predicted continuous reproduction of colors. Retro-cue trials produced better recall and higher rated attention. Critically, retro-cues improved recall to a similar extent across all levels of self-rated attention. This indicates that attention has multi-faceted and independent contributions to VWM.

Working Memory Swap Errors Have Identifiable Neural Representations

Working memory is an essential component of cognition that facilitates goal-directed behavior. Famously, it is severely limited and performance suffers when memory load exceeds an individual's capacity. Modeling of visual working memory responses has identified two likely types of errors: guesses and swaps. Swap errors may arise from a misbinding between the features of different items. Alternatively, these errors could arise from memory noise in the feature dimension used for cueing a to-be-tested memory item, resulting in the wrong item being selected. Finally, it is possible that so-called swap errors actually reflect informed guessing, which could occur at the time of a cue, or alternatively, at the time of the response. Here, we combined behavioral response modeling and fMRI pattern analysis to test the hypothesis that swap errors involve the active maintenance of an incorrect memory item. After the encoding of six spatial locations, a retro-cue indicated which location would be tested after memory retention. On accurate trials, we could reconstruct a memory representation of the cued location in both early visual cortex and intraparietal sulcus. On swap error trials identified with mixture modeling, we were able to reconstruct a representation of the swapped location, but not of the cued location, suggesting the maintenance of the incorrect memory item before response. Moreover, participants subjectively responded with some level of confidence, rather than complete guessing, on a majority of swap error trials. Together, these results suggest that swap errors are not mere response-phase guesses, but instead result from failures of selection in working memory, contextual binding errors, or informed guesses, which produce active maintenance of incorrect memory representations.

How Working Memory and Reinforcement Learning Are Intertwined: A Cognitive, Neural, and Computational Perspective

Reinforcement learning and working memory are two core processes of human cognition and are often considered cognitively, neuroscientifically, and algorithmically distinct. Here, we show that the brain networks that support them actually overlap significantly and that they are less distinct cognitive processes than often assumed. We review literature demonstrating the benefits of considering each process to explain properties of the other and highlight recent work investigating their more complex interactions. We discuss how future research in both computational and cognitive sciences can benefit from one another, suggesting that a key missing piece for artificial agents to learn to behave with more human-like efficiency is taking working memory's role in learning seriously. This review highlights the risks of neglecting the interplay between different processes when studying human behavior (in particular when considering individual differences). We emphasize the importance of investigating these dynamics to build a comprehensive understanding of human cognition.

Joint representation of working memory and uncertainty in human cortex

Neural representations of visual working memory (VWM) are noisy, and thus, decisions based on VWM are inevitably subject to uncertainty. However, the mechanisms by which the brain simultaneously represents the content and uncertainty of memory remain largely unknown. Here, inspired by the theory of probabilistic population codes, we test the hypothesis that the human brain represents an item maintained in VWM as a probability distribution over stimulus feature space, thereby capturing both its content and uncertainty. We used a neural generative model to decode probability distributions over memorized locations from fMRI activation patterns. We found that the mean of the probability distribution decoded from retinotopic cortical areas predicted memory reports on a trial-by-trial basis. Moreover, in several of the same mid-dorsal stream areas, the spread of the distribution predicted subjective trial-by-trial uncertainty judgments. These results provide evidence that VWM content and uncertainty are jointly represented by probabilistic neural codes.

Investigating the validity of the Perceptual Awareness Scale - The effect of task-related difficulty on subjective rating

The Perceptual Awareness Scale (PAS) is often used to probe conscious experience, but the assumptions behind the scale and its validity are rarely tested. Using a continuous colour estimation task to assess perceptual quality, we focus on how well PAS follows perceptual quality and how the presence of the estimation task affects PAS ratings. We varied the number and presentation time of stimuli in a simultaneous presentation and target position in a sequential presentation. In all experiments, PAS rating closely followed colour precision. However, it was affected by task-induced response bias, even when the possible task difficulty was judged, without performing the task itself. Still, this bias was only observed on the absolute rating level rather than the scale's ability to capture changes in perceptual quality. Reported studies shed light on factors influencing scales outside of the scale formulation and construction.

Uncertainty is maintained and used in working memory

What are the contents of working memory? In both behavioral and neural computational models, a working memory representation is typically described by a single number, namely, a point estimate of a stimulus. Here, we asked if people also maintain the uncertainty associated with a memory and if people use this uncertainty in subsequent decisions. We collected data in a two-condition orientation change detection task; while both conditions measured whether people used memory uncertainty, only one required maintaining it. For each condition, we compared an optimal Bayesian observer model, in which the observer uses an accurate representation of uncertainty in their decision, to one in which the observer does not. We find that this “Use Uncertainty” model fits better for all participants in both conditions. In the first condition, this result suggests that people use uncertainty optimally in a working memory task when that uncertainty information is available at the time of decision, confirming earlier results. Critically, the results of the second condition suggest that this uncertainty information was maintained in working memory. We test model variants and find that our conclusions do not depend on our assumptions about the observer's encoding process, inference process, or decision rule. Our results provide evidence that people have uncertainty that reflects their memory precision on an item-specific level, maintain this information over a working memory delay, and use it implicitly in a way consistent with an optimal observer. These results challenge existing computational models of working memory to update their frameworks to represent uncertainty.

The development of metacognitive accuracy in working memory across childhood

Growth in working memory capacity, the number of items kept active in mind, is thought to be an important aspect of childhood cognitive development. Here, we focused on participants' awareness of the contents of their working memory, or meta-working memory, which seems important because people can put cognitive abilities to best use only if they are aware of their limitations. In two experiments on the development of meta-working memory in children between 6 and 13 years old and adults, participants were to remember arrays of colored squares and to indicate if a probe item was in the array. On many trials, before the probe recognition test, they reported a metajudgment, how many items they thought they remembered. We compared meta-working memory judgments to actual performance and looked for associations between these measures on individual and trial-by-trial levels. Despite much lower working memory capacity in younger children there was little change in meta-working memory judgments across age groups. Consequently, younger participants were much less realistic in their metajudgments concerning their working memory capability. Higher cognitive capacity was associated with more accurate meta-working memory judgments within an age group. Trial-by-trial tuning of metajudgments was evident only in young adults and then only for small array set sizes. In sum, meta-working memory ability is a sophisticated skill that develops with age and may be an integral aspect of the development of working memory across the school years. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Allocation of resources in working memory: Theoretical and empirical implications for visual search

Recently, working memory (WM) has been conceptualized as a limited resource, distributed flexibly and strategically between an unlimited number of representations. In addition to improving the precision of representations in WM, the allocation of resources may also shape how these representations act as attentional templates to guide visual search. Here, we reviewed recent evidence in favor of this assumption and proposed three main principles that govern the relationship between WM resources and template-guided visual search. First, the allocation of resources to an attentional template has an effect on visual search, as it may improve the guidance of visual attention, facilitate target recognition, and/or protect the attentional template against interference. Second, the allocation of the largest amount of resources to a representation in WM is not sufficient to give this representation the status of attentional template and thus, the ability to guide visual search. Third, the representation obtaining the status of attentional template, whether at encoding or during maintenance, receives an amount of WM resources proportional to its relevance for visual search. Thus defined, the resource hypothesis of visual search constitutes a parsimonious and powerful framework, which provides new perspectives on previous debates and complements existing models of template-guided visual search.

Prediction of complex stimuli across saccades

The visual system can predict visual features across saccades based on learned transsaccadic associations between peripheral and foveal input. This has been shown for simple visual features such as shape, size, and spatial frequency. The present study investigated whether transsaccadic predictions are also made for more complex visual stimuli. In an acquisition phase, new transsaccadic associations were established. In the first experiment, pictures of real-world objects changed category during the saccade (fruits were changed into balls or vice versa). In the second experiment, the gender of faces was manipulated during the saccade (faces changed from male to female or vice versa). In the following test phase, the stimuli were briefly presented in the periphery, and participants had to indicate which object or face, respectively, they had perceived. In both experiments, peripheral perception was biased toward the acquired associated foveal input. These results demonstrate that transsaccadic predictions are not limited to a small set of simple visual features but can also be made for more complex and realistic stimuli. Multiple new associations can be learned within a short time frame, and the resulting predictions appear to be object specific.

Statistical learning as a reference point for memory distortions: Swap and shift errors

Humans use regularities in the environment to facilitate learning, often without awareness or intent. How might such regularities distort long-term memory? Here, participants studied and reported the colors of objects in a long-term memory paradigm, uninformed that certain colors were sampled more frequently overall. When participants misreported an object's color, these errors were often centered around the average studied color (i.e., "Rich" color), demonstrating swap errors in long-term memory due to imposed statistical regularities. We observed such swap errors regardless of memory load, explicit knowledge, or the distance in color space between the correct color of the tested object and the Rich color. An explicit guessing strategy where participants intentionally made swap errors when uncertain could not fully account for our results. We discuss other potential sources of observed swap errors such as false memory and implicit biased guessing. Although less robust than swap errors, evidence was also observed for subtle shift errors towards or away from the Rich color dependent on the color distance between the correct color and the Rich color. Together, these findings of swap and shift errors provide converging evidence for memory distortion mechanisms induced by a reference point, bridging a gap in the literature between how attention to regularities similarly influences visual working memory and visual long-term memory.

Humans incorporate trial-to-trial working memory uncertainty into rewarded decisions

Information stored in working memory (WM) is incorporated into many daily decisions and actions, and many complex decisions involve WM; however, there has been little work on investigating what WM information is used in memory decisions. Here we try to draw connections between WM and decision making by manipulating prior beliefs in a standard WM task with rewards. We use this paradigm to show that WM contains a representation of the trial-by-trial uncertainty of visual stimuli. This uncertainty is incorporated into rewarded decisions along with other information, such as expectations about the environment. By studying WM in parallel with decision making, we can gain new insight into how these systems work together.

Working memory (WM) plays an important role in action planning and decision making; however, both the informational content of memory and how that information is used in decisions remain poorly understood. To investigate this, we used a color WM task in which subjects viewed colored stimuli and reported both an estimate of a stimulus color and a measure of memory uncertainty, obtained through a rewarded decision. Reported memory uncertainty is correlated with memory error, showing that people incorporate their trial-to-trial memory quality into rewarded decisions. Moreover, memory uncertainty can be combined with other sources of information; after inducing expectations (prior beliefs) about stimuli probabilities, we found that estimates became shifted toward expected colors, with the shift increasing with reported uncertainty. The data are best fit by models in which people incorporate their trial-to-trial memory uncertainty with potential rewards and prior beliefs. Our results suggest that WM represents uncertainty information, and that this can be combined with prior beliefs. This highlights the potential complexity of WM representations and shows that rewarded decision can be a powerful tool for examining WM and informing and constraining theoretical, computational, and neurobiological models of memory.

An Electrophysiological Dissociation of Encoding vs. Maintenance Failures in Visual-Spatial Working Memory

Working memory (WM) performance varies substantially among individuals but the precise contribution of different WM component processes to these functional limits remains unclear. By analyzing different types of responses in a spatial WM task, we recently demonstrated a functional dissociation between confident and not-confident errors reflecting failures of WM encoding and maintenance, respectively. Here, we use event-related brain potentials to further explore this dissociation. Healthy participants performed a delayed orientation-discrimination task and rated their response confidence for each trial. The encoding-related N2pc component was significantly reduced for confident errors compared to confident correct responses, which is indicative of an encoding failure. In contrast, the maintenance-related contra-lateral delay activity was similar for these response types indicating that in confident error trials, WM representations – potentially the wrong ones – were maintained accurately and with stability throughout the delay interval. However, contra-lateral delay activity measured during the early part of the delay period was decreased for not-confident errors, potentially reflecting compromised maintenance processes. These electrophysiological findings contribute to a refined understanding of the encoding and maintenance processes that contribute to limitations in WM performance and capacity.

Theory of neural coding predicts an upper bound on estimates of memory variability

Observers reproducing elementary visual features from memory after a short delay produce errors consistent with the encoding-decoding properties of neural populations. While inspired by electrophysiological observations of sensory neurons in cortex, the population coding account of these errors is based on a mathematical idealization of neural response functions that abstracts away most of the heterogeneity and complexity of real neuronal populations. Here we examine a more physiologically grounded model based on the tuning of a large set of neurons recorded in macaque V1 and show that key predictions of the idealized model are preserved. Both models predict long-tailed distributions of error when memory resources are taxed, as observed empirically in behavioral experiments and commonly approximated with a mixture of normal and uniform error components. Specifically, for an idealized homogeneous neural population, the width of the fitted normal distribution cannot exceed the average tuning width of the component neurons, and this also holds to a good approximation for more biologically realistic populations. Examining eight published studies of orientation recall, we find a consistent pattern of results suggestive of a median tuning width of approximately 20°, which compares well with neurophysiological observations. The finding that estimates of variability obtained by the normal-plus-uniform mixture method are bounded from above leads us to reevaluate previous studies that interpreted a saturation in width of the normal component as evidence for fundamental limits on the precision of perception, working memory, and long-term memory.

A Metacognitive Perspective of Visual Working Memory With Rich Complex Objects

Visual working memory (VWM) has been extensively studied in the context of memory capacity. However, less research has been devoted to the metacognitive processes involved in VWM. Most metacognitive studies of VWM studies tested simple, impoverished stimuli, whereas outside of the laboratory setting, we typically interact with meaningful, complex objects. Thus, the present study aimed to explore the extent to which people are able to monitor VWM of real-world objects that are more ecologically valid and further afford less inter-trial interference. Specifically, in three experiments, participants viewed a set of either four or six memory items, consisting of images of unique real-world objects that were not repeated throughout the experiment. Following the memory array, participants were asked to indicate where the probe item appeared (Experiment 1) whether it appeared at all (Experiment 2) or whether it appeared and what was its temporal order (Experiment 3). VWM monitoring was assessed by subjective confidence judgments regarding participants’ objective performance. Similar to common metacognitive findings in other domains, we found that subjective judgments overestimated performance and underestimated errors, even for real-world, complex items held in VWM. These biases seem not to be task-specific as they were found in temporal, spatial, and identity VWM tasks. Yet, the results further showed that meaningful, real-world objects were better remembered than distorted items, and this memory advantage also translated to metacognitive measures.

Swap errors in spatial working memory are guesses

In typical visual working memory tasks, participants report the color of a previously studied item at some probed location. Alternatively, in some recent studies, a color is probed and participants must report the item's location. There is a surprising difference between these tasks: in location reports participants almost never guess randomly as they do when reporting color, but often incorrectly report the locations of non-probed items. This finding has been taken as evidence for feature binding errors in memory, and evidence against discrete capacity models, which predict that pure guessing should occur. We test an alternative possibility: that non-target responses are guesses, but intelligent ones. In particular, when asked to report the location of an item for which participants have no memory, they may guess near locations where they know something was presented. Here we present false-probe trials in which a color is probed that was not actually studied, and find that the responses, which are necessarily guesses, are nonetheless centered around studied locations. Moreover, we find that the confidence ratings for non-target responses are low, and similar to confidence for uniformly distributed guesses. In a second experiment, we find that manipulating the retention interval, which is known to affect guess rates, changes the rate of these low-confidence non-target responses. These results suggest that the tendency to report locations of non-probed items reflects a good guessing strategy; not something fundamental about how features and objects are represented in working memory.

Attentional capture alters feature perception

We live in a dynamic, distracting world. When distracting information captures attention, what are the consequences for perception? Previous literature has focused on effects such as reaction time (RT) slowing, accuracy decrements, and oculomotor capture by distractors. In the current study, we asked whether attentional capture by distractors can also more fundamentally alter target feature representations, and if so, whether participants are aware of such errors. Using a continuous report task and novel confidence range report paradigm, we discovered 2 types of feature-binding errors when a distractor was presented along with the target: First, when attention is strongly captured by the distractor, participants commit swapping errors (misreporting the color at the distractor location instead of the target color), which remarkably seem to occur without awareness. Second, when participants successfully resist capture, they tend to exhibit repulsion (perceptual distortion away from the color at the distractor location). Thus, we found that capture not only induces a spatial shift of attention, it also alters feature perception in striking ways. (PsycINFO Database Record (c) 2019 APA, all rights reserved).

Differential trajectories of memory quality and guessing across sequential reports from working memory

Working memory enables the storage of few items for a short period of time. Previous research has shown that items in working memory cannot be accessed equally well, indicating that they are held in at least two different states with different capacity limitations. However, it is unclear whether differences between states are due to limitations of the number of items that can be stored, or the quality with which items are stored. We employed a sequential whole-report procedure where participants reported the remembered orientation of each of two or four encoded Gabor patches. In addition, they rated their memory confidence prior to each report. Participants performed 600 trials per condition, allowing us to obtain reliable subjective ratings and estimates of precision, guessing, and misreport using a mixture model, separately for each sequential report. Different measures of memory quality consistently showed discontinuous trajectories across reports with a steep drop from the first to the second remembered item but only slight decreases thereafter. In contrast, both reported and modeled guessing changed continuously across reports. Our results support the notion of two states in working memory and show that they are distinguished by memory quality rather than quantity.

Visual expectations change subjective experience without changing performance

It is widely believed that visual expectations can change the subjective experiences of humans. We investigated how visual expectations in a recognition task affected objective performance and subjective perception. Using a 2-alternative-forced-choice task based on digit recognition of briefly presented and visually masked digits, we found over two experiments that expectations changed the quality of the experiences without changing the performance capabilities associated with the quality of experience. Expectations were manipulated by providing a cue indicating the set of possible digits that might appear on each trial. The results also inform the debate about whether subjective experiences can be categorized in a dichotomous manner or in a graded manner. We found that subjective experiences were graded near the objective threshold and more dichotomous away from the threshold. Furthermore, distinct expectations resulted in a more dichotomous distribution of subjective experience. We also provide evidence of an interesting relationship between stimulus duration, objective performance and subjective ratings. Only experiences that were rated as evoking some degree of perception showed systematic improvements in objective performance as a function of stimulus duration. These findings suggest that subjective experience cannot be understood without considering the broader cognitive context, namely that the quality of subjective experiences is dependent on a multitude of factors such as attention, task requirements and cognitive expectations.

Visual short-term memory through the lifespan: Preserved benefits of context and metacognition

Visual short-term memory (VSTM) ability falls throughout the life span in healthy adults. Using a continuous report task, in a large, population-based sample, we first confirmed that this decline affects the quality and quantity of reported memories as well as knowledge of which item went where. Visual and sensorimotor precision also worsened with advancing age, but this did not account for the reduced memory performance. We then considered two strategies that older individuals might be able to adopt, to offset these memory declines: the use of contextual encoding, and metacognitive monitoring of performance. Context and metacognitive awareness were both associated with significantly better performance, however these effects did not interact with age in our sample. This suggests that older adults retain their capacity to boost memory performance through attention to external context and monitoring of their performance. Strategies that focus on taking advantage of these preserved abilities may therefore help to maintain VSTM performance with advancing age. The article reports on analysis of the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data. (PsycINFO Database Record

Chunking as a rational strategy for lossy data compression in visual working memory

The nature of capacity limits for visual working memory has been the subject of an intense debate that has relied on models that assume items are encoded independently. Here we propose that instead, similar features are jointly encoded through a "chunking" process to optimize performance on visual working memory tasks. We show that such chunking can: (a) facilitate performance improvements for abstract capacity-limited systems, (b) be optimized through reinforcement, (c) be implemented by center-surround dynamics, and (d) increase effective storage capacity at the expense of recall precision. Human performance on a variant of a canonical working memory task demonstrated performance advantages, precision detriments, interitem dependencies, and trial-to-trial behavioral adjustments diagnostic of performance optimization through center-surround chunking. Models incorporating center-surround chunking provided a better quantitative description of human performance in our study as well as in a meta-analytic dataset, and apparent differences in working memory capacity across individuals were attributable to individual differences in the implementation of chunking. Our results reveal a normative rationale for center-surround connectivity in working memory circuitry, call for reevaluation of memory performance differences that have previously been attributed to differences in capacity, and support a more nuanced view of visual working memory capacity limitations: strategic tradeoff between storage capacity and memory precision through chunking contribute to flexible capacity limitations that include both discrete and continuous aspects. (PsycINFO Database Record

Separating memoranda in depth increases visual working memory performance

Visual working memory is the mechanism supporting the continued maintenance of information after sensory inputs are removed. Although the capacity of visual working memory is limited, memoranda that are spaced farther apart on a 2-D display are easier to remember, potentially because neural representations are more distinct within retinotopically organized areas of visual cortex during memory encoding, maintenance, or retrieval. The impact on memory of spatial separability in depth is less clear, even though depth information is essential to guiding interactions with objects in the environment. On one account, separating memoranda in depth may facilitate performance if interference between items is reduced. However, depth information must be inferred indirectly from the 2-D retinal image, and less is known about how visual cortex represents depth. Thus, an alternative possibility is that separation in depth does not attenuate between-items interference; it may even impair performance, as attention must be distributed across a larger volume of 3-D space. We tested these alternatives using a stereo display while participants remembered the colors of stimuli presented either near or far in the 2-D plane or in depth. Increasing separation in-plane and in depth both enhanced performance. Furthermore, participants who were better able to utilize stereo depth cues showed larger benefits when memoranda were separated in depth, particularly for large memory arrays. The observation that spatial separation in the inferred 3-D structure of the environment improves memory performance, as is the case in 2-D environments, suggests that separating memoranda in depth might reduce neural competition by utilizing cortically separable resources.

Strategic allocation of working memory resource

Visual working memory (VWM), the brief retention of past visual information, supports a range of cognitive functions. One of the defining, and largely studied, characteristics of VWM is how resource-limited it is, raising questions about how this resource is shared or split across memoranda. Since objects are rarely equally important in the real world, we ask how people split this resource in settings where objects have different levels of importance. In a psychophysical experiment, participants remembered the location of four targets with different probabilities of being tested after a delay. We then measured their memory accuracy of one of the targets. We found that participants allocated more resource to memoranda with higher priority, but underallocated resource to high- and overallocated to low-priority targets relative to the true probability of being tested. These results are well explained by a computational model in which resource is allocated to minimize expected estimation error. We replicated this finding in a second experiment in which participants bet on their memory fidelity after making the location estimate. The results of this experiment show that people have access to and utilize the quality of their memory when making decisions. Furthermore, people again allocate resource in a way that minimizes memory errors, even in a context in which an alternative strategy was incentivized. Our study not only shows that people are allocating resource according to behavioral relevance, but suggests that they are doing so with the aim of maximizing memory accuracy.

A resource-rational theory of set size effects in human visual working memory

Encoding precision in visual working memory decreases with the number of encoded items. Here, we propose a normative theory for such set size effects: the brain minimizes a weighted sum of an error-based behavioral cost and a neural encoding cost. We construct a model from this theory and find that it predicts set size effects. Notably, these effects are mediated by probing probability, which aligns with previous empirical findings. The model accounts well for effects of both set size and probing probability on encoding precision in nine delayed-estimation experiments. Moreover, we find support for the prediction that the total amount of invested resource can vary non-monotonically with set size. Finally, we show that it is sometimes optimal to encode only a subset or even none of the relevant items in a task. Our findings raise the possibility that cognitive "limitations" arise from rational cost minimization rather than from constraints.

Correlated individual differences suggest a common mechanism underlying metacognition in visual perception and visual short-term memory

Adaptive behaviour depends on the ability to introspect accurately about one's own performance. Whether this metacognitive ability is supported by the same mechanisms across different tasks is unclear. We investigated the relationship between metacognition of visual perception and metacognition of visual short-term memory (VSTM). Experiments 1 and 2 required subjects to estimate the perceived or remembered orientation of a grating stimulus and rate their confidence. We observed strong positive correlations between individual differences in metacognitive accuracy between the two tasks. This relationship was not accounted for by individual differences in task performance or average confidence, and was present across two different metrics of metacognition and in both experiments. A model-based analysis of data from a third experiment showed that a cross-domain correlation only emerged when both tasks shared the same task-relevant stimulus feature. That is, metacognition for perception and VSTM were correlated when both tasks required orientation judgements, but not when the perceptual task was switched to require contrast judgements. In contrast with previous results comparing perception and long-term memory, which have largely provided evidence for domain-specific metacognitive processes, the current findings suggest that metacognition of visual perception and VSTM is supported by a domain-general metacognitive architecture, but only when both domains share the same task-relevant stimulus feature.

Evidence of gradual loss of precision for simple features and complex objects in visual working memory

Previous studies have suggested that people can maintain prioritized items in visual working memory for many seconds, with negligible loss of information over time. Such findings imply that working memory representations are robust to the potential contaminating effects of internal noise. However, once visual information is encoded into working memory, one might expect it to inevitably begin degrading over time, as this actively maintained information is no longer tethered to the original perceptual input. Here, we examined this issue by evaluating working memory for single central presentations of an oriented grating, color patch, or face stimulus, across a range of delay periods (1, 3, 6, or 12 s). We applied a mixture-model analysis to distinguish changes in memory precision over time from changes in the frequency of outlier responses that resemble random guesses. For all 3 types of stimuli, participants exhibited a clear and consistent decline in the precision of working memory as a function of temporal delay, as well as a modest increase in guessing-related responses for colored patches and face stimuli. We observed a similar loss of precision over time while controlling for temporal distinctiveness. Our results demonstrate that visual working memory is far from lossless: while basic visual features and complex objects can be maintained in a quite stable manner over time, these representations are still subject to noise accumulation and complete termination. (PsycINFO Database Record

Confident failures: Lapses of working memory reveal a metacognitive blind spot

Working memory performance fluctuates dramatically from trial to trial. On many trials, performance is no better than chance. Here, we assessed participants' awareness of working memory failures. We used a whole-report visual working memory task to quantify both trial-by-trial performance and trial-by-trial subjective ratings of inattention to the task. In Experiment 1 (N = 41), participants were probed for task-unrelated thoughts immediately following 20% of trials. In Experiment 2 (N = 30), participants gave a rating of their attentional state following 25% of trials. Finally, in Experiments 3a (N = 44) and 3b (N = 34), participants reported confidence of every response using a simple mouse-click judgment. Attention-state ratings and off-task thoughts predicted the number of items correctly identified on each trial, replicating previous findings that subjective measures of attention state predict working memory performance. However, participants correctly identified failures on only around 28% of failure trials. Across experiments, participants' metacognitive judgments reliably predicted variation in working memory performance but consistently and severely underestimated the extent of failures. Further, individual differences in metacognitive accuracy correlated with overall working memory performance, suggesting that metacognitive monitoring may be key to working memory success.

Informed guessing in change detection

Provided stimuli are highly distinct, the detection of changes between two briefly separated arrays appears to be achieved by an all-or-none process where either the relevant information is in working memory or observers guess. This observation suggests that it is possible to estimate the average number of items an observer was able to retain across a series of trials, a potentially highly informative cognitive characteristic. For each version of the change detection paradigm, for this estimate to be accurate, it is important to specify how observers use the information available to them. For some instantiations of this task it is possible that observers use knowledge of the contents of working memory even when they are in a guessing state, rather than selecting between the response alternatives at random. Here we test the suggestion that observers may be able to use their knowledge of the number of items in memory to guide guessing in two versions of the change detection task. The four experiments reported here suggest that participants are, in fact, able to use the parameters of the task to update their base expectation of a change occurring to arrive at more informed guessing. (PsycINFO Database Record

Accounting for stimulus-specific variation in precision reveals a discrete capacity limit in visual working memory

If we view a visual scene that contains many objects, then momentarily close our eyes, some details persist while others seem to fade. Discrete models of visual working memory (VWM) assume that only a few items can be actively maintained in memory, beyond which pure guessing will emerge. Alternatively, continuous resource models assume that all items in a visual scene can be stored with some precision. Distinguishing between these competing models is challenging, however, as resource models that allow for stochastically variable precision (across items and trials) can produce error distributions that resemble random guessing behavior. Here, we evaluated the hypothesis that a major source of variability in VWM performance arises from systematic variation in precision across the stimuli themselves; such stimulus-specific variability can be incorporated into both discrete-capacity and variable-precision resource models. Participants viewed multiple oriented gratings, and then reported the orientation of a cued grating from memory. When modeling the overall distribution of VWM errors, we found that the variable-precision resource model outperformed the discrete model. However, VWM errors revealed a pronounced "oblique effect," with larger errors for oblique than cardinal orientations. After this source of variability was incorporated into both models, we found that the discrete model provided a better account of VWM errors. Our results demonstrate that variable precision across the stimulus space can lead to an unwarranted advantage for resource models that assume stochastically variable precision. When these deterministic sources are adequately modeled, human working memory performance reveals evidence of a discrete capacity limit. (PsycINFO Database Record

Monkeys and humans take local uncertainty into account when localizing a change

Since sensory measurements are noisy, an observer is rarely certain about the identity of a stimulus. In visual perception tasks, observers generally take their uncertainty about a stimulus into account when doing so helps task performance. Whether the same holds in visual working memory tasks is largely unknown. Ten human and two monkey subjects localized a single change in orientation between a sample display containing three ellipses and a test display containing two ellipses. To manipulate uncertainty, we varied the reliability of orientation information by making each ellipse more or less elongated (two levels); reliability was independent across the stimuli. In both species, a variable-precision encoding model equipped with an “uncertainty–indifferent” decision rule, which uses only the noisy memories, fitted the data poorly. In both species, a much better fit was provided by a model in which the observer also takes the levels of reliability-driven uncertainty associated with the memories into account. In particular, a measured change in a low-reliability stimulus was given lower weight than the same change in a high-reliability stimulus. We did not find strong evidence that observers took reliability-independent variations in uncertainty into account. Our results illustrate the importance of studying the decision stage in comparison tasks and provide further evidence for evolutionary continuity of working memory systems between monkeys and humans.

Clear evidence for item limits in visual working memory

There is a consensus that visual working memory (WM) resources are sharply limited, but debate persists regarding the simple question of whether there is a limit to the total number of items that can be stored concurrently. Zhang and Luck (2008) advanced this debate with an analytic procedure that provided strong evidence for random guessing responses, but their findings can also be described by models that deny guessing while asserting a high prevalence of low precision memories. Here, we used a whole report memory procedure in which subjects reported all items in each trial and indicated whether they were guessing with each response. Critically, this procedure allowed us to measure memory performance for all items in each trial. When subjects were asked to remember 6 items, the response error distributions for about 3 out of the 6 items were best fit by a parameter-free guessing model (i.e. a uniform distribution). In addition, subjects' self-reports of guessing precisely tracked the guessing rate estimated with a mixture model. Control experiments determined that guessing behavior was not due to output interference, and that there was still a high prevalence of guessing when subjects were instructed not to guess. Our novel approach yielded evidence that guesses, not low-precision representations, best explain limitations in working memory. These guesses also corroborate a capacity-limited working memory system - we found evidence that subjects are able to report non-zero information for only 3-4 items. Thus, WM capacity is constrained by an item limit that precludes the storage of more than 3-4 individuated feature values.

Superior Intraparietal Sulcus Controls the Variability of Visual Working Memory Precision

Limitations of working memory (WM) capacity depend strongly on the cognitive resources that are available for maintaining WM contents in an activated state. Increasing the number of items to be maintained in WM was shown to reduce the precision of WM and to increase the variability of WM precision over time. Although WM precision was recently associated with neural codes particularly in early sensory cortex, we have so far no understanding of the neural bases underlying the variability of WM precision, and how WM precision is preserved under high load. To fill this gap, we combined human fMRI with computational modeling of behavioral performance in a delayed color-estimation WM task. Behavioral results replicate a reduction of WM precision and an increase of precision variability under high loads (5 > 3 > 1 colors). Load-dependent BOLD signals in primary visual cortex (V1) and superior intraparietal sulcus (IPS), measured during the WM task at 2-4 s after sample onset, were modulated by individual differences in load-related changes in the variability of WM precision. Although stronger load-related BOLD increase in superior IPS was related to lower increases in precision variability, thus stabilizing WM performance, the reverse was observed for V1. Finally, the detrimental effect of load on behavioral precision and precision variability was accompanied by a load-related decline in the accuracy of decoding the memory stimuli (colors) from left superior IPS. We suggest that the superior IPS may contribute to stabilizing visual WM performance by reducing the variability of memory precision in the face of higher load.

SIGNIFICANCE STATEMENT

This study investigates the neural bases of capacity limitations in visual working memory by combining fMRI with cognitive modeling of behavioral performance, in human participants. It provides evidence that the superior intraparietal sulcus (IPS) is a critical brain region that influences the variability of visual working memory precision between and within individuals (Fougnie et al., 2012; van den Berg et al., 2012) under increased memory load, possibly in cooperation with perceptual systems of the occipital cortex. These findings substantially extend our understanding of the nature of capacity limitations in visual working memory and their neural bases. Our work underlines the importance of integrating cognitive modeling with univariate and multivariate methods in fMRI research, thus improving our knowledge of brain-behavior relationships.

Fechner's law in metacognition: A quantitative model of visual working memory confidence

Although visual working memory (VWM) has been studied extensively, it is unknown how people form confidence judgments about their memories. Peirce (1878) speculated that Fechner's law-which states that sensation is proportional to the logarithm of stimulus intensity-might apply to confidence reports. Based on this idea, we hypothesize that humans map the precision of their VWM contents to a confidence rating through Fechner's law. We incorporate this hypothesis into the best available model of VWM encoding and fit it to data from a delayed-estimation experiment. The model provides an excellent account of human confidence rating distributions as well as the relation between performance and confidence. Moreover, the best-fitting mapping in a model with a highly flexible mapping closely resembles the logarithmic mapping, suggesting that no alternative mapping exists that accounts better for the data than Fechner's law. We propose a neural implementation of the model and find that this model also fits the behavioral data well. Furthermore, we find that jointly fitting memory errors and confidence ratings boosts the power to distinguish previously proposed VWM encoding models by a factor of 5.99 compared to fitting only memory errors. Finally, we show that Fechner's law also accounts for metacognitive judgments in a word recognition memory task, which is a first indication that it may be a general law in metacognition. Our work presents the first model to jointly account for errors and confidence ratings in VWM and could lay the groundwork for understanding the computational mechanisms of metacognition. (PsycINFO Database Record

The role of long-term memory in a test of visual working memory: Proactive facilitation but no proactive interference

We report 4 experiments examining whether associations in visual working memory are subject to proactive interference from long-term memory (LTM). Following a long-term learning phase in which participants learned the colors of 120 unique objects, a working memory (WM) test was administered in which participants recalled the precise colors of 3 concrete objects in an array. Each array in the WM test consisted of 1 old (previously learned) object with a new color (old-mismatch), 1 old object with its old color (old-match), and 1 new object. Experiments 1 to 3 showed that WM performance was better in the old-match condition than in the new condition, reflecting a beneficial contribution from LTM. In the old-mismatch condition, participants sometimes reported colors associated with the relevant shape in LTM, but the probability of successful recall was equivalent to that in the new condition. Thus, information from LTM only intruded in the absence of reportable information in WM. Experiment 4 tested for, and failed to find, proactive interference from the preceding trial in the WM test: Performance in the old-mismatch condition, presenting an object from the preceding trial with a new color, was equal to performance with new objects. Experiment 5 showed that long-term memory for object-color associations is subject to proactive interference. We conclude that the exchange of information between LTM and WM appears to be controlled by a gating mechanism that protects the contents of WM from proactive interference but admits LTM information when it is useful. (PsycINFO Database Record

Large Capacity of Conscious Access for Incidental Memories in Natural Scenes

When searching a crowd, people can detect a target face only by direct fixation and attention. Once the target is found, it is consciously experienced and remembered, but what is the perceptual fate of the fixated nontarget faces? Whereas introspection suggests that one may remember nontargets, previous studies have proposed that almost no memory should be retained. Using a gaze-contingent paradigm, we asked subjects to visually search for a target face within a crowded natural scene and then tested their memory for nontarget faces, as well as their confidence in those memories. Subjects remembered up to seven fixated, nontarget faces with more than 70% accuracy. Memory accuracy was correlated with trial-by-trial confidence ratings, which implies that the memory was consciously maintained and accessed. When the search scene was inverted, no more than three nontarget faces were remembered. These findings imply that incidental memory for faces, such as those recalled by eyewitnesses, is more reliable than is usually assumed.