Elsevier

Cognitive Psychology

Volume 97, September 2017, Pages 79-97
Cognitive Psychology

Clear evidence for item limits in visual working memory

https://doi.org/10.1016/j.cogpsych.2017.07.001Get rights and content

Highlights

  • In a whole report task, participants report all items held in Working Memory (WM).

  • We directly observed variable precision in working memory responses.

  • For large arrays, many responses were best fit by a uniform guessing distribution.

  • Robust evidence for guessing confirms the presence of item limits in WM storage.

Abstract

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.

Introduction

Working memory (WM) is an online memory system where information is maintained in the service of ongoing cognitive tasks. Although there is a broad consensus that WM resources are sharply limited, there has been sustained debate about the precise nature of these limits. On the one hand, discrete-resource models argue that only a handful of items can be maintained at one time, such that some items fail to be stored when the number of memoranda exceeds the observer’s capacity (Awh et al., 2007, Cowan, 2001, Rouder et al., 2008, Zhang and Luck, 2008). On the other hand, continuous resource models argue that WM storage depends on a central pool of resources that can be divided across an unlimited number of items (Bays and Husain, 2008, van den Berg et al., 2012, Wilken and Ma, 2004).

Of course, it has long been known that memory performance declines as the number of memoranda increases in a WM task. For example, Luck and Vogel (1997) varied the number of simple colors in a change detection task that required subjects to detect whether one of the memoranda had changed between two sequential presentations of a sample display. They found that while performance was near ceiling for set sizes up to three items, accuracy declined quickly as set size increased beyond that point. This empirical pattern is well described by a model in which subjects store 3–4 items in memory and then fail to store additional items. However, the same data can be accounted for by a continuous resource model that presumes storage of all items, but with declining precision as the number of memoranda increases (Wilken & Ma, 2004). According to continuous resource accounts, increased errors with larger set sizes are caused by insufficient mnemonic precision rather than by storage failures (but for a critique of this account see Nosofsky & Donkin, 2016b). Thus, a crux issue in this literature has been to distinguish whether performance declines with displays containing more than a handful of items are due to storage failures or sharp reductions in mnemonic precision.

In this context, Zhang and Luck (2008) offered a major step forward with an analytic approach that provides separate estimates of the probability of storage and the quality of the stored representations. They employed a continuous recall WM task in which subjects were cued to recall the precise color of an item from a display with varying numbers of memoranda. Their key insight was that if subjects failed to store a subset of the items, there should be two qualitatively distinct types of responses within a single distribution of response errors. If subjects had stored the probed item in memory, responses should be centered on the correct color, with a declining frequency of responses as the distance from the correct answer increased. But if subjects had failed to store the probed item, then responses should be random with respect to the correct answer, producing a uniform distribution of answers across the entire space of possible colors. Indeed, their data revealed that the aggregate response error distribution was well described as a weighted average of target-related and guessing responses. Thus, Zhang and Luck (2008) provided some of the first positive evidence that working memory performance reflects a combination of target-related and guessing responses.

Subsequent work, however, has argued that the empirical pattern reported by Zhang and Luck (2008) can be explained by continuous resource models that presume storage of all items in every display (van den Berg et al., 2014, van den Berg et al., 2012). A key feature of these models has been the assumption that precision in visual WM may vary substantially; thus, while some items may be represented precisely, other representations in memory may contain little information about the target item. Using this assumption, van den Berg et al. (2012) showed that they could account for the full distribution of errors – including apparent guessing – and that their model outperformed the one proposed by Zhang and Luck (2008). Indeed, converging evidence from numerous studies has left little doubt that precision varies across items in these tasks (e.g., Fougnie, Suchow, & Alvarez, 2012). That said, the question of whether precision is variable is logically separate from the question of whether observers ever fail to store items in these procedures. To examine the specific reasons why one model might achieve a superior fit over another, it is necessary to explore how distinct modeling decisions influence the outcome of the competition. Embracing this perspective, van den Berg et al. (2014) carried out a factorial comparison of WM models in which the presence of items limits and the variability of precision were independently assessed. Although this analysis provided clear evidence that mnemonic precision varies across items and trials, the data were not decisive regarding the issue of whether working memory is subject to an item limit. There was a numerical advantage for models that endorsed item limits, but it was not large enough to draw strong conclusions. Thus, the critical question of whether item limits in visual working memory elicit guessing behavior remains unresolved.

Here, we report data that offer stronger traction regarding this fundamental question about the nature of limits in working memory. Much previous work has focused on explaining variance within aggregate response-error distributions (i.e. the shape of the response distribution and how it changes across set sizes). Here, we chose a different route. Rather than developing a new model that might explain a small amount of additional variance in “traditional” partial report datasets, we developed a new experimental paradigm in which subjects recalled—in any order that they wished—the precise color (Experiment 1a) or orientation (Experiment 1b) of every item in the display. This procedure has the key benefit of measuring the quality of all simultaneously remembered items, and it yields the clear prediction that if there are no item limits, then there should be measurable information across all responses. To anticipate the results, this whole report procedure provided rich information about the quality of all items within a given trial as well as subjects’ metaknowledge of variations in quality. Observers consistently reported the most precisely remembered items first, yielding monotonic declines in information about the recalled item with each successive response. Critically, for the plurality of subjects, the final three responses made were best modeled by the parameter-free uniform distribution that indicates guessing. In additional analyses and experiments, we showed that subjective guess ratings tracked mixture model guessing parameter (Experiments 1 & 2), that output interference could not explain our estimates of capacity (Experiment 2), and that making subjective guess ratings did not drive our evidence for guessing (Experiment 3). Finally, we used simulations to question a key claim of the variable precision model – that representations used by this model all contain measurable information. Previously, others have suggested that the variable precision model may mimic guess responses with ultra-low precision representations (Nosofsky and Donkin, 2016b, Sewell et al., 2014). Here, we advanced these claims by showing that variable precision models that eschew guessing posit a high prevalence of memories that are indistinguishable from guesses. Moreover, the frequency of these putative representations precisely tracked the estimated rate of guessing in models that acknowledge item limits.

In sum, there has been a longstanding debate over whether there is any limit in the number of items that can be stored in working memory. Our findings provide compelling evidence that working memory is indeed subject to item limits, disconfirming a range of prior models that deny guessing entirely or posit an item limit that varies from trial to trial without any hard limit in the total number of items that can be stored (e.g. Sims et al., 2012, van den Berg et al., 2014). Instead, our results point toward a model where each individual has a capacity ceiling (e.g. 3 items), but they frequently under-achieve their maximum capacity, likely due to fluctuations in attentional control (Adam, Mance, Fukuda, & Vogel, 2015).

Section snippets

Subjects

22 subjects from the University of Oregon completed Experiment 1a for payment ($8/h) or class credit. All participants had normal color vision and normal or corrected-to-normal visual acuity, and all gave informed consent according to procedures approved by the University of Oregon institutional review board.

Stimuli

Stimuli were generated in MATLAB (The MathWorks, Inc., Natick, MA, www.mathworks.com) using the Psychophysics Toolbox extension (Brainard, 1997, Pelli, 1997). Stimuli were presented on a

Subjects

17 subjects from the University of Oregon completed Experiment 2a. All participants had normal color vision and normal or corrected-to-normal visual acuity, and they were compensated with payment ($8/h) or course credit. All participants gave informed consent according to procedures approved by the University of Oregon institutional review board.

Procedures

Stimuli and procedures in Experiments 2a were identical to Experiment 1a except for the order in which responses were collected. On each trial, the

Materials & methods

Ten participants from the University of Chicago community participated in Experiment 3 for payment ($10/h). Stimuli and procedures were nearly identical to Experiment 1b; participants were asked to remember the orientation of 1, 2, 3, 4 or 6 items and to report all items in any order they chose. However, participants did not make confidence ratings using the two mouse buttons. During the instruction period, the experimenter instructed the participants that they should remember all of the items.

Simulation results: variable precision models mimic guessing behaviors by positing very low precision memories

Across several experiments, we have shown that guessing accounts for a large proportion of subjects’ responses when working memory load is high (e.g. ∼50% of Set Size 6 responses). This result is apparently in conflict with past findings that a variable precision model that denies guessing (hereafter called “VP-no guessing”) is a strong competitor for models that presume a high prevalence of storage failures (van den Berg et al., 2014). How does the VP-no guessing model achieve close fits of

General discussion

Understanding the nature of capacity limits in working memory has been a longstanding goal in memory research. The capacity debate has been dominated by a theoretical dichotomy of “number” versus “precision.” Discrete resource models have argued that capacity is limited by the number of items that can be concurrently stored, and that subjects to resort to guessing when more than a handful of memoranda are presented (Fukuda et al., 2010, Zhang and Luck, 2008). By contrast, continuous resource

Contributions

K.A. collected data and performed analyses. K.A., E.V., and E.A. conceived of experiments and wrote the manuscript.

Conflicts of interest

None.

Acknowledgements

We thank Irida Mance for collecting pilot data that were not included in the publication. Research was supported by National Institute of Mental Health grant 5R01 MH087214-08 and Office of Naval Research grant N00014-12-1-0972. Datasets for all experiments are available online on Open Science Framework at https://osf.io/kjpnk/.

References (45)

  • G.-Y. Bae et al.

    Stimulus-specific variability in color working memory with delayed estimation

    Journal of Vision

    (2014)
  • P.M. Bays et al.

    The precision of visual working memory is set by allocation of a shared resource

    Journal of Vision

    (2009)
  • P.M. Bays et al.

    Dynamic shifts of limited working memory resources in human vision

    Science

    (2008)
  • P. Berens

    CircStat: A MATLAB toolbox for circular statistics

    Journal of Statistical Software

    (2009)
  • D.H. Brainard

    The psychophysics toolbox

    Spatial Vision

    (1997)
  • N. Cowan

    The magical number 4 in short-term memory: A reconsideration of mental storage capacity

    The Behavioral and Brain Sciences

    (2001)
  • N. Cowan et al.

    Detection of the number of changes in a display in working memory

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2016)
  • C. Donkin et al.

    Verbal labeling, gradual decay, and sudden death in visual short-term memory

    Psychonomic Bulletin & Review

    (2015)
  • R.W. Engle et al.

    Individual differences in working memory capacity and what they tell us about controlled attention, general fluid intelligence, and functions of the prefrontal cortex

  • D. Fougnie et al.

    Variability in the quality of visual working memory

    Nature Communications

    (2012)
  • K. Fukuda et al.

    Quantity, not quality: The relationship between fluid intelligence and working memory capacity

    Psychonomic Bulletin & Review

    (2010)
  • J.M. Gold et al.

    Reduced capacity but spared precision and maintenance of working memory representations in schizophrenia

    Archives of General Psychiatry

    (2010)
  • Cited by (0)

    View full text