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

Download PDF

1608 Accesses
4 Citations
Explore all metrics

Abstract

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

Object-based selection in visual working memory

Article Open access 13 July 2021

Repulsion bias is insensitive to spatial attention, yet expands during active working memory maintenance

Article 11 June 2024

Retro-dimension-cue benefit in visual working memory

Article Open access 24 October 2016

Use our pre-submission checklist

Avoid common mistakes on your manuscript.

Introduction

Working memory (WM) is used to maintain and manipulate items in mind while using them to accomplish a goal (Baddeley, 2012). Many theorists propose that, in addition to actively retaining goal-relevant information, WM also involves actively deleting information that is no longer relevant^{Footnote 1} (Hasher et al., 2007; Lewis-Peacock et al., 2018; Oberauer, 2009). However, the results reported here suggest that this mechanism may not be universally applied.

Measuring the prioritization and deletion of items in working memory (WM) with retrocue tasks

While maintaining goal-relevant information in WM, it helps to be cued to the information that is most relevant for ongoing cognition (e.g., which item will be tested by an upcoming recall or recognition test) even if the cue comes after the information has been presented and encoded in WM (i.e., retrocues) (Souza & Oberauer, 2016). Numerous studies have shown differences in behavior (accuracy, response times) and brain activity (EEG/ERP, MEG, fMRI) associated with WM for retrocued versus uncued items (for a meta-analysis, see Wallis et al., 2015). Many theorists have concluded that these differences arise when internal attention selects and protects retrocued items against interference from other items in WM (Souza et al., 2016). Other theorists have posited that retrocueing benefits relevant items because controlled attentional processes can be strategically used to actively delete irrelevant items from WM (e.g., Lewis-Peacock et al., 2018). However, it is unclear whether cueing is beneficial because people selectively attend to and enhance the representation of relevant items, because they selectively delete irrelevant items, or because of both processes (Lintz & Johnson, 2021). Tasks with multiple retrocues are particularly revealing because the consequences of prioritizing one item over the rest can be assessed for both the initially cued and the initially uncued items.

Consider a situation in which two to-be-remembered items are presented and actively retained in WM; an initial retrocue indicates which item is to be tested first, and then a second retrocue indicates which item is to be tested second. Using such a task with categorical stimuli (faces, words, or directions of motion) as memoranda, Rose et al. (2016) showed behavioral and neural evidence that supported the idea that no-longer-relevant items were actively deleted from WM. When the two items were initially presented and retained in WM, the category of both items could be decoded from the participant’s brain activity using fMRI or EEG. Following the first retrocue, neural representation of the uncued item dropped to baseline as if it were no longer actively retained “in WM” – but it could be reactivated by a single pulse of transcranial magnetic stimulation (TMS) applied to a category-selective region of posterior cortex. This suggested that, while this uncued item was still potentially relevant later on in the trial, the uncued item was passively retained in WM via “activity-silent” (short-term synaptic plasticity) mechanisms (Rose, 2020; Silvanto, 2017).^{Footnote 2} However, following the second retrocue, which indicated that the uncued item was no longer relevant on the trial, TMS could no longer reactivate the uncued (no-longer-relevant) item. This suggested that the no-longer-relevant item was actively deleted from WM following the second retrocue. For replications and extensions, see Fulvio and Postle (2020) and Wolff et al. (2017).

Related research has also shown evidence for an active-deletion process that removes items cued as no longer relevant for WM (Oberauer, 2018; for a review, see Lewis-Peacock et al., 2018). However, other evidence suggests that this active-deletion mechanism is not always utilized (Dagry et al., 2017; Dagry & Barrouillet, 2017; Lilienthal et al., 2015; Lintz & Johnson, 2021; Oberauer, 2018). Although the active-deletion hypothesis proposes that slower presentation rates allow more time to remove distractors and also that deleted items should be less accessible on subsequent memory tests, contradictory evidence has been shown from repetition priming, lexical decision, and subsequent memory effects of distractors (e.g., Dagry et al., 2017; Dagry & Barrouillet, 2017; Lilienthal et al., 2015), even when participants are explicitly instructed to either remove uncued items or refresh cued items (Lintz & Johnson, 2021).

Revealing the activation state and interference among items in WM with repulsion and attraction effects

Another way to test how items are retained in (or deleted from) WM is to examine the extent to which retained items interfere with one another during recall (Wildegger et al., 2015). For stimuli that are represented in a continuous feature space such as orientations, spatial locations, colors, etc., it is possible to detect subtle numerical biases between items retained in WM (Bae & Luck, 2017). For example, when attempting to recall a cued item in WM (e.g., an orientation of 30°), an uncued item in WM (e.g., an orientation of 10°) can systematically bias recall of the cued item either away from the uncued item – a phenomenon called “repulsion” (Kiyonaga & Egner, 2016) – or toward the uncued item – a phenomenon called “attraction” (Chunharas et al., 2022). In this example, repulsion would be reflected by the participant recalling the target orientation to be farther in the feature space from the distractor (e.g., 32° vs. 28°). Such biases have also been shown to influence WM for color (Golomb, 2015), motion (Czoschke et al., 2019), and even faces whose features vary along continua (Mallett et al., 2020). Bias can come from both task-irrelevant distractors or memoranda from previous trials, as in the so-called serial dependence effect (Shan & Postle, 2022). The phenomenon is consistent with neurocomputational models of visual WM that posit repulsive bias between similar items due to lateral inhibition (Johnson et al., 2009), which can flip in sign from an attractive bias from a previous trial to a repulsive bias within a trial (Fritsche et al., 2020).

The present study

The present study used orientations and a similar double-retrocue paradigm to that used by Rose et al. (2016) to examine repulsion and attraction effects of competing memory items (Fig. 1).

In a double-retrocue task, bias should be largest on recall 1 (when the uncued item is still potentially relevant on the trial). If the uncued item is actively deleted after the second cue (because it is no longer relevant on the trial), then bias should be smaller on recall 2 than on recall 1. However, larger bias from the uncued item on recall 2 compared to recall 1 would suggest that the no-longer-relevant item persists in WM; such findings would question the generalizability of the active-deletion mechanism.

Our task design and analyses allowed the measurement of memory fidelity of items held in such cued or uncued states, as well as the relative contributions of distinct sources of influence on their recall. The use of continuous stimuli enabled us to explore the effects of dropping an uncued memory item from an attended state on memory, which is more sensitive to detecting subtle biases and the sources of variability in recall than other paradigms such as recognition of categorical stimuli. Specifically, we used computational modeling to separate memory errors into precision (defined as the standard deviation of errors), guess rate (defined as the likelihood that the participant had no memory trace for the target item), and swap error rate (defined as the likelihood a response reflects the uncued memory item; also known as a binding error) (Peters et al., 2019). We predicted that precision would be worse and guess and swap error rates would be higher for recall-2-switch trials, in which the initially uncued item was cued for recall on the second test, compared to recall-1 and recall-2-stay trials. Bias analyses were examined with the mixture model results in an attempt to elucidate the source of differences between cued and uncued items on these parameters. We had no a priori hypotheses about the direction of bias (repulsion or attraction) or differences between the conditions. The main hypothesis regarding bias was that if no-longer-relevant items were actively deleted, then bias should be less on recall 2 than on recall 1 responses. Any contrary evidence would call for a revision to the active-deletion hypothesis.

Experiment 1

Method

Participants

Forty-one (M_age = 19.1 years, range = 18–35 years, 27 female) right-handed students with normal or corrected-to-normal vision were recruited to participate in the experiment. Participants provided informed consent (Institutional Review Board (IRB) protocol 17-02-3629) and were remunerated with cash or course credit (US$15 or 1 credit/h). Data for six participants were unavailable (three withdrew after screening, three due to technical errors); analyses were conducted on the remaining 35 participants’ data.^{Footnote 3}

WM task

Participants were seated approximately 37 cm (n = 19) or 57 cm (n = 16) away from a 24-in. ASUS computer monitor with 1,920 × 1,080 resolution and a 60-Hz refresh rate. The task and stimuli were generated and run in MATLAB using the Psychophysics Toolbox V3.0 (Brainard, 1997; Kleiner et al., 2007; Pelli, 1997). Responses were given using the “1” and “2” buttons on the T9 number pad of a standard QWERTY keyboard to freely rotate the presented recall stimulus counterclockwise or clockwise, respectively.

Stimulus details

Central fixation was identified by a white circle with an outer radius of ${}^{{\raisebox{.5ex}{1}\!\left/ \!\raisebox{-.5ex}{4}\right.}}$ pixel and an inner radius of ${}^{{\raisebox{.5ex}{1}\!\left/ \!\raisebox{-.5ex}{8}\right.}}$ pixel. The experimental stimuli consisted of two sine-wave gratings (i.e., Gabor patches) with a diameter of 2°, spatial frequency of 2 cycles/°, a phase of 0, and a Michelson contrast of 100%. The orientations were separated into seven distinct orientation bins with centers of 13°, 39°, 65°, 91°, 117°, 143°, and 169°, with each bin containing the same number of orientations. For a given trial, orientations were selected pseudo-randomly from these bins with a jitter of ± 5°, and the two stimuli in a given trial varied by more than 10°.

Location of stimulus presentation in the lower left and right visual hemifields was matched to phosphene localizations acquired from participants in an ongoing TMS study to target early visual cortex (V1/V2). The retrocues consisted of circular outlines surrounding the locations where the stimuli were presented. The cued item was outlined by a bold (0.5°) white circle; the non-cued item was outlined by a non-bold (0.15)° light-grey circle. Presenting circles at the locations of both the cued and uncued items was necessary to avoid selectively “pinging” the cued item with a visual impulse (Wolff et al., 2017).

Phosphene localization procedure

The locations where stimuli were presented was determined by an ongoing rTMS study in which a different set of participants first underwent a phosphene localization and thresholding procedure in a dark room to determine if they could reliably see a circular-shaped phosphene in the lower-right visual field when holding central-fixation from single-pulses of TMS applied to left, early-visual cortex (V1/V2). If so, the TMS intensity at which a phosphene was induced in five out of ten trials was determined following established procedures (Abrahamyan et al., 2011; Rademaker et al., 2017). Then TMS intensity was set to 110% of the phosphene threshold, single pulses were applied at the localized area, and, following each pulse, participants were instructed to use the computer mouse to trace an outline of the perceived phosphene onto the black computer screen with a gray, central-fixation cross using custom MATLAB/PsychToolbox code.

Following the drawing of at least ten outlines, each outline was fit to an ellipse using the fitellipse function, the centroid of each ellipse was calculated, and the median centroid value (in X and Y screen-pixel coordinates) was recorded. These coordinates were used to determine the location at which the center of the right Gabor-orientation-patch was presented for the WM task. The left Gabor patch was presented in the contralateral visual field from these coordinates; the right Gabor patch was presented in the mirroring side of the visual field. Therefore, stimuli locations were individually determined and unique for each participant in the ongoing TMS study. For the purposes of this behavioral-only, control experiment, different participants were randomly matched to the stimuli-locations determined for participants who completed the phosphene localization task and the TMS version of the experiment in order to assess the potential impact of stimulus-location variability on performance. That is, participants in this behavioral-only, control study did not receive TMS; the locations at which stimuli were presented for a participant were matched to those that were generated for a corresponding participant in the TMS experiment.^{Footnote 4}

Task procedure

A white fixation circle was presented on a black screen at the beginning of each trial for 2 s and remained on the screen throughout stimulus and cue presentation (Fig. 1). Gabor patches were presented for 0.2 s in the lower visual-hemifield, one in the right-hemifield and the other symmetrically mirrored in the left-hemifield according to the locations determined by the phosphene-localization procedure described in the preceding section. After a 2-s delay, the first retrocue was presented for 0.5 s. A 2.5-s delay followed the cue before a random Gabor orientation was presented in the center of the screen. Participants were instructed to rotate the orientation to match the orientation of the cued-stimulus. Once the response was submitted, feedback was displayed at central-fixation for 0.3 s. Responses that were within 15°, 15–30°, or >30° away from the target turned the fixation cross green, yellow, or red, respectively.^{Footnote 5} Following the first feedback, a second retrocue was displayed for 0.5 s. This retrocue could signal that either the same stimulus would be tested a second time (a “stay” trial) or that the originally uncued stimulus would be tested (a “switch” trial). Trials were balanced so that there was an equal number of stay and switch trials in each block. A random Gabor patch was once again presented in the center of the screen after the 2.5-s delay, and participants rotated the Gabor patch to match the stimulus cued by the second retrocue. Feedback was once again given following the second recall response. Each block consisted of 56 trials, and participants completed 2-3 blocks in each session.

Data quality checks

For the average accuracy analysis, in order to identify potential outliers in the data, for each recall condition per participant, the errors were first converted to z-scores, and any z-score > 3 or < −3 was removed from the data set. Since these responses were significant outliers, they likely reflect cases in which participants had no memory representation for the target item and resorted to guessing. Therefore, removing these responses before the average accuracy analysis enabled us to get a more accurate measure of memory performance. A total of 1.3% of the responses was removed (212 out of a total of 15,770 responses), and no more than 14 trials were removed from any recall condition for an individual participant. The analysis of errors was conducted on the non-z-score converted data as the circular deviation of the recalled orientation from the target orientation in degrees. The boxplot function in R Studio was then used across all participants to determine any outliers in the dataset (defined as 1.5 times the interquartile range above or below the third or first quartiles, respectively); two participants were determined to be outliers and removed from the error analysis comparing behavioral performance across recall conditions. Thus, data from 33 participants were used in the behavioral analyses.

For the mixture model analyses, all trials for the remaining participants were included (even the trials previously identified by the z-score analysis as outliers) because the models attempted to separate errors by different parameters, so were able to account for outliers. One participant had an implausible recall-2-switch precision parameter (3.27E+28), suggesting that the mixture model failed to fit the data. Therefore, parameter values for this participant were not included in the group level analysis, leaving data from 32 participants to be included in the mixture model analyses. A mixed-design ANOVA showed that the interaction between performance on the three recall conditions and viewing distance was not significant (F(2,62) = 0.71, p = 0.50). Moreover, the correlations between performance and degrees of visual angle were not significant for any of the three recall conditions, rs = 0.05, 0.04, and 0.05, respectively, ps > 0.25. For the nontarget bias analysis, we included all trials, but removed the data of the two participants previously deemed outliers as in the average accuracy analysis.

Data analysis

R Studio was used to perform all of the statistical tests on the model parameters and comparisons. The normality of the distribution of recall errors for each condition was assessed using a log10 transformation and Shapiro-Wilks tests, which confirmed normality (ps > 0.1442, see Fig. 2). For all analyses, two-tailed, Bonferroni-corrected t-tests were used, unless stated otherwise.

Errors were calculated as the circular difference between the target orientation and the response orientation, according to the von Mises distribution. The difference between the nontarget orientation (i.e., the uncued orientation) and the response was also calculated for the mixture modeling to determine the influence that the nontarget had on the response.

Mixture model analyses

Mixture modeling was conducted on the errors using MATLAB and the MemToolbox (Suchow et al., 2013). The models were used to parameterize memory precision and the proportion of responses in which the participant likely guessed or committed a binding error. We plotted the response errors centered around the target response of 0 error. The Standard Mixture Model (Zhang & Luck, 2008) was compared to the Swap Model (Bays et al., 2009). The Standard Mixture Model used the distance of a response from the target value to determine both the probability that the error reflects the precision (reflected by SD) of the participant’s memory for the target item and the probability that the response was a random guess (reflected by the uniform distribution called the guess rate, or g parameter). This model uses the following equation when fitting the data:

$$p\left(\hat{\theta}\right)=\left(1-\gamma \right){\phi}_{\sigma}\left(\hat{\theta}-\theta \right)+\gamma \frac{1}{2\pi }$$

(1)

where 𝜃 serves as the target value (in radians), $\hat{\theta}$ serves as the response value, 𝛾 serves as the frequency of random guesses, and 𝜙_𝜎 serves as the circular analogue of the von Mises distribution (mean = 0, SD = 𝜎).

The Swap Model (Bays et al., 2009) includes the same precision and guess rate parameters as well as a third parameter, the swap error rate, which reflects the probability that a response reflects a memory for the nontarget item. By taking the accuracy of the response relative to the nontarget item into account, the swap error rate indicates the probability that the participant recalled the uncued item rather than the cued item. The Swap Model is described by the equation:

$$p\left(\hat{\theta}\right)=\left(1-\gamma -\beta \right){\phi}_{\sigma}\left(\hat{\theta}-\theta \right)+\gamma \frac{1}{2\pi }+\beta \frac{1}{m}\sum_i^m{\phi}_{\sigma}\left(\hat{\theta}-{\theta}_i^{\ast}\right)$$

(2)

where 𝛽 serves as the probability of a swap error and {𝜃₁*,𝜃₂*,...𝜃_m*} are the m nontarget line orientation values (Bays et al., 2009).

The responses for each recall condition (recall-1, recall-2-stay, and recall-2-switch) were modeled separately for each participant to see how memory changed when items were switched from an unprioritized- to a prioritized-state within-subjects. The fits of each model were compared using the Akaike Information Criterion (AIC).

Nontarget bias analyses

To assess the influence of the nontarget (uncued) orientation on recall of the target orientation we calculated and compared the amount of response bias relative to the nontarget on each of the three recall conditions for both the average amount of bias and as a function of the degree of difference between the target and nontarget orientations. Positive or negative error value indicates whether the response was biased away from the nontarget (“repulsion”) or toward the nontarget (“attraction”), respectively. We performed Bonferroni-corrected t-tests on the average error relative to the nontarget across all trials to compare the amount of bias between the three recall conditions (using paired-sample t-tests) and for the average of bins of trials with small, medium, or large differences between the target and nontarget orientations (using one-sample t-tests vs. zero, i.e., no bias).

Results

The distributions of recall errors relative to the to-be-remembered target orientation on recall-1, recall-2-stay, and recall-2-switch trials for all participants are shown in Fig. 2.

To determine the consequences of holding information in an unprioritized state, we compared the average accuracies across the three recall conditions (recall-1, recall-2-stay, and recall-2-switch trials) as the absolute value of recall error (in degrees). Recall error was higher on recall-2-switch trials (M = 22.7, SD = 8.06) than on both recall-1 (M = 14.2, SD = 5.19; t(32) = -14.68, p < 0.001) and recall-2-stay trials (M = 14.6, SD = 5.77; t(32) = -13.50, p < 0.001), but there was no difference between recall-1 and recall-2-stay trials (t(32) = -0.69, p = 0.50) (OSM Fig. 1). These results support our hypotheses that shifting a memory item into an unattended state weakened the fidelity of memory for that item compared to items maintained in an attended state. We then conducted mixture model analyses on the data in order to better understand the source(s) of the differences.

Mixture modeling

Model preference

We first compared the two mixture models to determine which of the models was a better fit to the data. We performed a Wilcoxon signed-rank test on the difference in the AIC values between the Standard Mixture Model and the Swap Model for all participants (as in Bays & Taylor, 2018). The Swap Model was preferred over the Standard Mixture Model for all three recall conditions: recall-1 (ΔM = 15.33, p < 0.001); recall-2-stay (ΔM = 6.77, p < 0.01); and recall-2-switch (ΔM = 9.45, p < 0.01).

Parameter differences

To determine the effects of shifting attention between items, we compared the error parameters from the Swap Model across the three recall conditions (recall-1, recall-2-stay, and recall-2-switch). Three Bonferroni-corrected, two-tailed (unless stated otherwise) paired t-tests were performed for each parameter to compare all three recall conditions.

Precision

As predicted, there was a statistically significant increase in the precision parameter (indicating less precision) on recall-2-switch trials compared with both recall-1 trials (t(31) = -5.64, p < 0.01) and recall-2-stay trials (t(31) = -5.61, p < 0.01). There was no significant difference in the precision parameter between recall-1 trials and recall-2-stay trials (t(31) = -0.70, p = 0.49, Fig. 3A).

Guess rate

The guess rate was higher for recall-2-switch trials than for both recall-1 trials (t(31) = -7.34, p < 0.001) and recall-2-stay trials (t(31) = -5.26, p < 0.001), and there was no significant difference between recall-1 and recall-2-stay trials (t(31) = -1.94, p = 0.06, Fig. 3B).

Swap error rate

As predicted, the estimated swap error rate was higher on recall-2-switch trials than on both recall-1 and recall-2-stay trials (ts(31) = -6.25 and -4.83, ps < 0.01 and .017, respectively);^{Footnote 6} the difference in swap error rate between recall-1 and recall-2-stay trials did not survive Bonferroni correction (t(31) = -2.18, p = 0.04, Fig. 3B).

Overall, these data support our hypotheses that holding an item in a deprioritized state results in worse memory fidelity for that item and also increases the commission of swap errors. This also increased the number of guesses.

Nontarget bias analysis: repulsion and attraction effects

To elucidate the source of the differences between the recall conditions we investigated the role that the nontarget played in biasing response errors and how this bias changed across the recall conditions. We performed Bonferroni-corrected paired t-tests on the average error with bias across the three recall conditions. In contrast to the active-deletion hypothesis, there was no difference in repulsive bias between recall-1 and recall-2-stay trials (t(32) = 0.15, p = 0.88), and there was less repulsive bias on recall-1 than recall-2-switch trials (t(32) = 3.52, p < 0.005). There was also more repulsive bias on recall-2-switch versus stay trials despite the fact that the uncued item was no longer relevant during recall 2 for both stay and switch trials (t(32) = -3.00, p < 0.01, Fig. 4).

To further elucidate the source of bias on recall, we calculated the amount of bias as a function of the difference between the target and nontarget orientations. The purpose of this analysis was to assess whether the amount of bias that the uncued (nontarget) item had on recall of the cued (target) item depended on the similarity between the target and nontarget. Trials were binned around three orientation differences centered around relatively small (~25°), medium (~50°), and large (~75°) differences between the target and nontarget orientations.^{Footnote 7} The amount of bias on recall-1 and recall-2-stay trials was not significantly different from zero for the 25°, 50°, or 75° bins (ps > 0.05). For recall-2-switch trial, there was significant repulsion from the nontarget orientation when there were small or medium differences between the target and nontarget (ps < 0.01), but the amount of bias was not significant when there were large (~75°) differences between the target and the non-target (p = 0.69, Fig. 5). As discussed below, current neurocomputational models of visual WM posit that lateral inhibition mechanisms could drive repulsive bias seen between similar items (Johnson et al., 2009).

Are differences in precision, guess, and swap parameters due to bias?

Finally, an exploratory correlational analysis was done to see if the poorer precision, guess, and swap error parameters on recall-2-switch trials that were observed (Fig. 3) were associated with the increase in repulsive bias that was observed on these trials (Figs. 4 and 5). Participants’ precision parameter and their average bias on recall-2-switch trials were positively correlated (r = 0.41, p = 0.02), indicating that those with poorer memory precision had greater repulsive bias; in contrast, participants’ guess and swap parameters were not correlated with their average amount of bias (rs = 0.04 and -0.29, ps = 0.83 and 0.11), indicating that the increase in guess and swap errors was not associated with the increase in bias on recall-2-switch trials. Also note that recall of nontargets (swap errors) would result in an attractive bias – not the observed repulsive bias that differed depending the degree of target-nontarget similarity.

Discussion

Compared to actively maintaining and recalling a cued item in WM, passively retaining and then returning an uncued item back into focal attention resulted in decreases in recall precision (which was associated with the degree of bias from the nontarget orientation), and increases in the probability that the participant guessed or recalled the nontarget item. These findings are consistent with hypotheses that internal attention can select one of multiple items in WM to prioritize its retention and recall over other items, and that items dropped from focal attention can be passively retained and reactivated when needed, via error-prone retrieval processes (see also LaRocque et al., 2015; Peters et al., 2019).

The key finding is that, contrary to the hypothesis that no-longer-relevant items are actively deleted from WM, these items persisted and biased recall of the target item held in focal attention, especially when the target and no-longer-relevant items were similar to one another. Moreover, recall-1 trials showed the same amount of bias as recall-2-stay trials and less bias than recall-2-switch trials, which contradicts the pattern predicted by the active-deletion hypothesis. Following the second retrocue, the uncued item was no longer relevant and, therefore, according to the active-deletion hypothesis, it should have been deleted from WM and should have resulted in less bias for responses on recall-2-stay and recall-2-switch trials. The present results suggest that no-longer-relevant items were not deleted from WM following the second retrocue.

One potential explanation for this pattern of results is that, when trying to remember the two line orientations, participants may have encoded the two distinct orientations as one “chunked” representation. Participants could have bound the two distinct orientation objects into one chunked representation, with both orientations bound together as an angle or clock hands, for example. Anecdotal evidence from post-experimental debriefing of our participants is consistent with this interpretation. Although the two Gabor orientations were presented separately in the lower left and right hemifields, many participants reported encoding the two as a bound object, (e.g., an angle by projecting the lines out to their intersecting point, like the hour and minute hands on an analog clock). Encoding the two objects as a bound object would change the way the relevant and irrelevant features are represented. If two stimuli retained in WM are bound or “chunked” into a single object, then it may not be possible to fully delete the no-longer-relevant item from WM following a retrocue. This might explain the pattern of results, which differs from paradigms with retrieval of more discrete (e.g., categorical) stimuli that cannot be as easily bound into a single object, such as a face paired with either a word or a direction of motion, as in Rose et al. (2016) (see also Fulvio & Postle, 2020). To test this account of the biases from the no-longer-relevant stimulus a second experiment was conducted.

Experiment 2

Experiment 2 used the same task design as Experiment 1. The only difference was that we explicitly instructed participants to bind or “chunk” the two orientations together on each trial by mentally connecting the line orientations into one bound object. We told participants to imagine the line orientations’ point of intersection and think about the two orientations as an angle or hands of a clock. If the source of the bias from the no-longer-relevant item on recall of the target item that was observed in Experiment 1 was due to this binding at encoding, then the pattern of results should be similar for Experiment 2. If the pattern is not similar then the source of bias must be due to some other factor.