Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Working Memory Content Is Distorted by Its Use in Perceptual Comparisons

Working Memory Content Is Distorted by Its Use in Perceptual Comparisons Visual information around us is rarely static. To perform a task in such a dynamic environment, we often have to compare current visual input with our working memory (WM) representation of the immediate past. However, little is known about what happens to a WM representation when it is compared with perceptual input. To test this, we asked young adults (N = 170 total in three experiments) to compare a new visual input with a WM representation prior to reporting the WM representation. We found that the perceptual comparison biased the WM report, especially when the input was subjectively similar to the WM representation. Furthermore, using computational modeling and individual- differences analyses, we found that this similarity-induced memory bias was driven by representational integration, rather than incidental confusion, between the WM representation and subjectively similar input. Together, our findings highlight a novel source of WM distortion and suggest a general mechanism that determines how WM interacts with new visual input. Keywords working memory, memory distortion, individual differences, open data Received 10/16/20; Revision accepted 9/22/21 Many daily activities rely on our ability to maintain during maintenance (e.g., Bennett & Cortese, 1996; accurate mental representations of the immediate past Magnussen & Greenlee, 1992; Nemes et  al., 2012; Sun to guide our behavior. For example, consider a Good et  al., 2017; Teng & Kravitz, 2019) and at the time of Samaritan who witnesses a phone falling out of a pass- memory reporting (Makovski et  al., 2008; Wang et  al., erby’s pocket on a busy street. The Good Samaritan 2018). For example, in a previous study, participants were would look down to pick up the phone in hopes of asked to maintain a color in WM while completing a rapid returning it to its owner. After doing so, they would serial visual presentation (RSVP) task in which a stream need to find the owner again by comparing their mem- of letters was presented on a differing color patch (Teng ory of the owner’s appearance with the other people & Kravitz, 2019). The researchers found that the color they see on the street. Past studies have demonstrated representation maintained in WM was distorted by the that working memory (WM) can actively maintain a task-irrelevant color patch, especially when the two colors small amount of task-relevant information (e.g., the were physically similar to each other. Therefore, WM owner’s appearance) over a brief delay when the infor- seems vulnerable to perceptual interference. mation is not perceptually present (e.g., while looking However, the true nature of this interference remains down on the street to pick up the phone; Luck & Vogel, unknown. For example, it is unclear whether the mem- 2013; Ma et al., 2014). However, little is known as to ory distortion occurs if a WM representation remains whether WM remains intact as it is used for memory- guided perceptual comparisons. Corresponding Author: Although one might assume that WM remains intact as Keisuke Fukuda, University of Toronto Mississauga, Department of it is used, past studies have shown that WM can be inter- Psychology fered with by subsequent perceptual inputs introduced Email: keisuke.fukuda@utoronto.ca Psychological Science 33(5) 817 within the focus of internal attention. Given that the intervening RSVP task used by Teng and Kravitz (2019) Statement of Relevance drew attention away from WM maintenance, the mem- Imagine a person who witnesses a hit-and-run ory distortion may be contingent on disruption of active traffic accident. When the witness tries to identify WM maintenance (Makovski et al., 2008; Wang et al., the license plate number of the car involved, a 2018). It is also possible that the WM representation bus occludes their sight of the car. When the bus was not necessarily distorted but was occasionally passes, they need to find the car by comparing replaced by the intervening stimulus instead (i.e., swap the cars they see now with the memory of the car errors), making WM reports appear biased (Bays et al., that committed the accident. In doing so, they 2009). Lastly, because previous studies did not measure might assume that their memory of the car remains the perceived similarity between WM and intervening intact as it is compared with the cars they see stimuli, it is not known whether the interference was now. However, the present study shows that this dictated by the physical similarity or perceived similar- assumption is simply not valid: As a memory rep- ity between them. resentation of the immediate past (i.e., working We addressed these questions in the present study memory) is compared with new perceptual infor- by having participants remember a single visual feature mation, the memory report becomes systemati- (i.e., color or shape) in WM and compare it with new cally biased toward perceptually similar input. stimuli. Therefore, the WM had to be actively main- Thus, our findings reveal an overlooked vulner- tained so that it could be compared with new stimuli. ability in working memory and suggest that the In three experiments, we found that WM reports were witness may misreport the license plate of a simi- distorted, especially when participants perceived the lar, but different, car to the police. new stimuli as similar to the WM. Critically, we found that this similarity-induced memory bias was larger fol- lowing perceived similarity than perceived dissimilarity reported large effect sizes (Cohen’s d > 1.5). Because we of new stimuli, even after controlling for their physical used different stimuli, we anticipated a medium to large similarity. We further validated the systematic effect of effect size (Cohen’s d = 0.8). A power analysis (conducted subjective similarity by demonstrating that memory in G*Power Version 3.1; Faul et al., 2007) determined that biases induced by an initial similarity judgment were at least 15 samples would be necessary to detect such an “canceled out” by an additional judgment that biased effect with a statistical power of .8. We recruited 16 par- the WM in an opposing direction. Computational mod- ticipants (11 female; mean age = 23.2 years) for Experi- eling revealed that the similarity-induced memory bias ment 1a and 16 participants (10 female; mean age = 22.6 was better accounted for by representational integration years) for Experiment 1b. Participants provided informed between WM and perceptually similar input than by consent, and the study was approved by the University of probabilistic swap errors. Thus, our findings reveal an Toronto Research Ethics Board. overlooked window of vulnerability in WM and suggest a general mechanism that determines how a new per- Apparatus, stimuli, and procedure. Stimuli were gen- ceptual input interacts with WM (e.g., Kiyonaga et al., erated in MATLAB (The MathWorks, Natick, MA) using the 2017; Rademaker et al., 2019; Sun et al., 2017). Psychophysics Toolbox (Brainard, 1997) and were pre- sented at 60 Hz on an LCD monitor. Viewing distance was Experiments 1a and 1b approximately 60 cm. The shape stimulus set used in Experiment 1a contained 360 shapes whose circular conti- To examine the consequences of comparing WM with nuity was empirically validated (Li et al., 2020). The color perceptual inputs, we had participants remember a stimuli used in Experiment 1b comprised 360 equally simple stimulus (i.e., shape in Experiment 1a and color spaced color values sampled from Commission Internatio- in Experiment 1b) and compare it with a new percep- nale de l’Éclairage (CIE) L*a*b* space with a* centered at 20 tual input (Fig. 1). Subsequently, participants reported and b* centered at 38 with a radius of 60. L* was set to 70. their memory of the original stimulus as precisely as Participants completed multiple trials in the baseline possible. Here, we predicted that the WM report would and perceptual-comparison conditions (Fig. 1b). In the be biased toward the new input, especially when the baseline condition, participants were presented with new input was subjectively similar to the WM. one stimulus—either a shape (3.8° × 3.8°) or a colored circle (5.2° diameter) for 800 ms. After a 3,200-ms reten- Method tion interval, participants were presented with a circular stimulus wheel (15.4° diameter). For the shape task, Participants. Related experiments examining the mem- the stimulus wheel was composed of 18 equidistant ory bias for orientation memory (Rademaker et al., 2015) 818 Fukuda et al. Shape Space Color Space “Remember” “Recall” “Adjust” “Confidence?” 1 2 3 “Similar or Not?” “Remember” “Recall” “Adjust” “Confidence?” 1 2 3 800 ms 800 ms 1,600 ms Until Response Until Response Until Response + 800-ms Delay Fig. 1. Stimulus spaces and example trial procedures from Experiments 1a and 1b. The stimulus spaces used for Experiments 1a (shape task) and 1b (color task) are shown in (a). Trial procedures from the shape task (b) are shown separately for the baseline and perceptual- comparison conditions. In each condition, participants were first presented with a target shape that they attempted to remember across a brief maintenance interval. Following the maintenance interval, participants reported the target shape by rotating a line indicator to select from a continuous wheel. Participants then adjusted their selection, if desired, to best match their memory before submitting their report with a confidence rating. In the perceptual-comparison condition, participants also performed a perceptual comparison on a probe shape that was presented during the maintenance interval by indicating whether the probe shape was similar or dissimilar to the target shape. The color task was identical to the shape task except for the stimulus type and the stimulus wheel. shapes (20° apart in the shape space). For the color looked similar or dissimilar to the original stimulus by task, it was a color wheel composed of the 360 colors pressing a key. The probe stayed on the screen for 1,600 (1° per color). To report the remembered stimulus, they ms. The probe was randomly sampled from 16° to 105° first selected an item that best matched the original away from the original stimulus on each trial. After an stimulus from the stimulus wheel by rotating a line 800-ms delay following the probe presentation, partici- indicator using the left and right arrow keys on the pants reported the original memory item as in the base- keyboard. After the indicator pointed toward the desired line condition. Participants performed 12 experimental item, participants pressed the space bar to confirm the blocks, and each block contained 12 baseline and 36 selection. The selected item then appeared at the center perceptual-comparison trials. of the screen for optional refinements using the left and right arrow keys. When satisfied, participants indicated Analysis. For each trial, a signed response offset was their confidence in their report (1 = high confidence, calculated in relation to the probe. A positive sign indi- 2 = low confidence, 3 = no confidence) by pressing cated a response offset toward the probe (for the distri- one of three keys on the keyboard. No time limit was bution of raw response offsets, see the supplementary imposed on the memory report in order to emphasize material available at https://osf.io/79kbm/). For the base- accuracy. line trials, the sign of the response offset was randomly In the perceptual-comparison condition, a probe assigned. To quantify the magnitude of the bias, we com- stimulus was presented 800 ms after the offset of the puted the mean of the signed response offsets for each original stimulus, and participants indicated whether it condition. We focused on trials with high-confidence Perceptual Baseline Comparison Psychological Science 33(5) 819 memory reports (> 68% of trials, or > 83 trials, in all con- modified paradigm. More precisely, participants clicked ditions; for the same pattern of results with all trials a mouse to report WM and performed two-alternative included, see the supplementary material available at forced-choice judgments for perceptual comparisons https://osf.io/79kbm/). (Fig. 3). More importantly, considering that many To isolate the effect of perceived similarity on the behaviors rely on the maintenance of unbiased WM memory bias, we first identified the probe distances that representations, we examined whether the similarity- resulted in both similar and dissimilar judgments on induced memory bias can be corrected by biasing WM separate trials within subjects (ambivalent probe dis- in the opposite direction in a subsequent perceptual tances). For each ambivalent probe distance, we calcu- judgment. If the bias can be corrected by subsequent lated the mean bias magnitude following similar and perceptual judgments, then its magnitude should be dissimilar judgments separately. The mean bias magni- smaller when the similar probes in two consecutive tudes following similar and dissimilar judgments were judgments were sampled from opposing directions (i.e., then averaged across all ambivalent probe distances. This clockwise and counterclockwise) than from the same procedure isolated the effect of perceived probe similar- direction. ity (indicated by participants’ judgments) on the bias magnitude while equating the effect of physical probe Method similarity (determined by the sampling procedure). Participants. On the basis of the results of Experiment 1 (Cohen’s d > 0.73), we anticipated a medium to large Results effect size (Cohen’s d = 0.73). A power analysis (con- Both shape (Experiment 1a) and color (Experiment 1b) ducted in G*Power Version 3.1; Faul et  al., 2007) deter- memory reports exhibited an attraction bias (> 0°) mined that 22 samples were necessary to detect such an toward the probe when the probe was judged to be effect with a statistical power of .9. After providing written similar to the memory item—shape: M = 7.61°, t(15) = informed consent to protocols approved by the University 4.70, p < .001, 95% confidence interval (CI) = [4.28°, of Toronto Research Ethics Board, 32 individuals (24 10.94°], Cohen’s d = 1.18 (Figs. 2a and 2b); color: M = female; mean age = 19.3 years) participated. The data from 8.14°, t(15) = 5.47, p < .001, 95% CI = [4.97°, 11.32°], four participants were removed because they did not com- Cohen’s d = 1.37 (Figs. 2e and 2f ). We also found some plete the experiment (n = 1), failed to follow instructions evidence for an attraction bias when the probe was (n = 1), or too infrequently made high-confidence mem- judged to be dissimilar to the original memory item— ory reports (i.e., < 15% for the baseline conditions; n = 2). shape: M = 1.49°, t(15) = 1.48, p = .159, 95% CI = [–0.65°, As a result, 28 participants’ data were analyzed. 3.64°], Cohen’s d = 0.37; color: M = 3.03°, t(15) = 2.93, p = .01, 95% CI = [0.83°, 5.12°], Cohen’s d = 0.73—but Procedure. The procedure was similar to that of Experi- the bias magnitude was significantly smaller than when ment 1. Participants performed four blocks of the color the probe was judged to be similar—similar shape ver- task and four blocks of the shape task in a pseudoran- sus dissimilar shape: M = 5.42°, t(15) = 3.14, p = .007, domized order. In the baseline conditions (Fig. 3a), par- 95% CI = [1.74°, 9.10°], Cohen’s d = 0.78; similar color ticipants remembered the original stimulus over a 500-ms versus dissimilar color: M = 5.11°, t(15) = 2.98, p = .009, (short delay) or a 5,500-ms (long delay) interval. Two 95% CI = [1.45°, 8.77°], Cohen’s d = 0.74. The trials with delays were introduced to establish the baseline response- ambivalent probe distances (shape task: mean range of offset distributions for the perceptual-comparison condi- probe distance = 20.1°–97.7°; color task: mean range of tions (for an example of the effect of delay on WM probe distance = 19.3°–97.6°) produced larger attraction precision, see Rademaker et  al., 2018). After the delay, biases when the probes were perceived to be similar participants used a mouse to click on an item that best rather than dissimilar to the original memory items— matched the original stimulus, fine-tuned their response shape task: ΔM = 6.19°, t(15) = 3.01, p = .009, 95% CI = using key presses, and then reported their confidence as [1.81°, 10.58°], Cohen’s d = 0.75 (Figs. 2c and 2d); color in Experiment 1. task: ΔM = 10.92°, t(15) = 4.50, p < .001, 95% CI = [5.75°, In the perceptual-comparison conditions, partici- 16.08°], Cohen’s d = 1.13 (Figs. 2g and 2h). pants performed two intervening perceptual compari- sons 500 ms after the original stimulus offset. Because subjective similarity of a given perceptual probe varied Experiment 2 across trials, we controlled the subjective similarity of Experiment 1 demonstrated that WM reports were a physically similar probe (e.g., 30° away from the biased when WM was directly compared with a subjec- target) by presenting it together with a physically dis- tively similar probe. In Experiment 2, we sought to similar probe (180° away from the similar probe) and replicate this similarity-induced memory bias in a had participants identify the more similar probe from HC Bias (°) for HC Bias (°) for Ambivalent Probe Trials Ambivalent Probe Trials ab cd 0.03 0.03 40 40 Baseline Similar 0.02 0.02 Dissimilar 20 20 0.01 0.01 0 0 −20 −20 −180 −90 0 90 180 −180 −90 0 90 180 Signed Response Offset for HC Reports (°) Signed Response Offset for HC Reports (°) ef gh 0.03 0.03 40 40 Baseline Similar 0.02 0.02 Dissimilar 20 20 0.01 0.01 0 0 0 0 −20 −20 0 90 180 0 90 180 −180 −90 −180 −90 Signed Response Offset for HC Reports (°) Signed Response Offset for HC Reports (°) Fig. 2. Results of Experiments 1a (top row) and 1b (bottom row). Probability distributions of signed response offsets for high-confidence (HC) reports (a, e) are shown for the baseline condition and for similar and dissimilar probe judgments in the perceptual-comparison condition. For demonstration purposes, the response proportion for a given signed response offset was computed by calculating the mean response proportion across a 30° window centered around the signed response offset. Positive offsets indicate memory bias toward the first similar probe. The broken lines indicate within-subjects standard errors of the mean. Bias magnitude for HC memory reports (b, f) is shown for similar (pink) and dissimilar (blue) probe judgments. In each violin plot, the thick horizontal line indicates the mean across participants, and the width of the violin indicates the density of the data. Positive values indicate memory bias toward the probe. Probability distributions of signed response offsets for HC reports for ambivalent probe trials (c, g) are shown for the baseline condition and for similar and dissimilar probe judgments in the perceptual-comparison condition. Bias magnitude for HC memory reports for ambivalent probe trials (d, h) is shown for similar (pink) and dissimilar (blue) probe judgments. In each violin plot, the thick horizontal line indicates the mean across participants, and the width of the violin indicates the density of the data. Positive values indicate memory bias toward the probe. Experiment 1b (Color) Experiment 1a (Shape) Response Proportion Response Proportion HC Bias (°) HC Bias (°) Response Proportion for Response Proportion for Ambivalent Probe Trials Ambivalent Probe Trials 821 “Remember” “Recall” “Adjust” “Confidence?” Confidence Report 1: High Confidence Short Delay 2: Low Confidence 3: No Confidence 1 2 3 Until Response Until Response Until Response “Remember” “Recall” “Adjust” “Confidence?” Long Delay 1 2 3 “Remember” “More Similar?” “More Similar?” “Recall” “Adjust” “Confidence?” Same-Side Probe 1 2 3 “Remember” “More Similar?” “Recall” “Adjust” “More Similar?” “Confidence?” Opposite-Side Probe 1 2 3 1,500 ms 500 ms 2,000 ms 2,000 ms Until Response Until Response Until Response + 500-ms Delay + 500-ms Delay Same-Side Probe Opposite-Side Probe Memory Item Memory Item 1st Probe Pair 2nd Probe Pair Fig. 3. Trial procedures and probe-sampling procedures for Experiment 2. Trial procedures from the color task (a) are shown separately for the baseline (short delay, long delay) and experimental (same-side probe, opposite-side probe) conditions. In each condition, participants were first presented with a target color that they attempted to remember across a brief maintenance interval. Following the maintenance interval, participants reported the target color by clicking on a continuous wheel. Participants then adjusted their selection, if desired, to best match their memory before submitting their report with a confidence rating. In the experimental conditions, participants also performed two consecutive perceptual comparisons during the maintenance interval. In each comparison, participants selected which of two simultaneously presented probe colors was more similar to the target color. The sampling of the similar probes is depicted in (b). In the same-side-probe condition, the similar probes were both sampled from the same side of the circular color space relative to the target. In the opposite-side-probe condition, the similar probes were sampled from opposite sides of the target. The dissimilar probes were sampled 180° from the similar probes in both pairs. The shape task was identical to the color task except for the stimulus type and the stimulus wheel. 822 Fukuda et al. the pair (Fig. 3). In each comparison, two probes were 7.25°], Cohen’s d = 1.36 (Fig. 4b); color task: M = 6.11°, presented on each side of the screen (5.2° from the t(27) = 6.30, p < .001, 95% CI = [4.12°, 8.09°], Cohen’s center of the screen), and they reported which probe d = 1.19 (Fig. 4d)—indicating that the memory reports looked more similar to the original stimulus by pressing were attracted toward the similar probes. In contrast, either the left or right arrow key. One of the probes the mean signed response offset for the opposite-side- was randomly sampled from ±16° to 45° away from the probe condition exhibited a nonsignificant negative original stimulus (i.e., similar probe). The other probe signed offset for shape, M = −0.36°, t(27) = −0.78, p = was sampled 180° away from the similar probe (i.e., .441, 95% CI = [–1.32°, 0.59°], Cohen’s d = −0.15, and dissimilar probe). The two probes remained on the a small but significant negative signed offset for color, screen for 2,000 ms regardless of the report. After the M = −1.29°, t(27) = 2.20, p = .037, 95% CI = [–2.49°, offset of the first pair of probes, a 500-ms delay fol- –0.09°], Cohen’s d = −0.42, indicating that the memory lowed, and the second pair of probes was presented reports were, if anything, biased away from the first for another similarity judgment (2,000 ms). After another similar probe. The magnitude of the bias (i.e., absolute 500-ms delay, participants reported the original stimulus values) for the same-side-probe condition was statisti- in the same manner as the baseline conditions. cally greater than the magnitude of the bias for the The two perceptual-comparison conditions differed opposite-side-probe conditions—shape task: ΔM = in how the similar probes were sampled (Fig. 3b). In 5.27°, t(27) = 7.46, p < .001, 95% CI = [3.82°, 6.72°], the same-side-probe condition, similar probes in each Cohen’s d = 1.41; color task: ΔM = 4.82°, t(27) = 3.88, pair were sampled from the same side of the stimulus p < .001, 95% CI = [2.27°, 7.37°], Cohen’s d = 0.73. space relative to the memory item. In the opposite-side- probe condition, similar probes in each pair were sam- Modeling of the Similarity-Induced pled from opposite sides of the stimulus space relative Memory Bias to the memory item. Participants completed 40 trials for each condition in a pseudorandomized order. Experiment 2 not only replicated the similarity-induced memory bias in a modified paradigm but also demon- Analysis. For each trial, a signed response offset was cal- strated that it could be corrected by an additional simi- culated as the response offset in the direction of the similar larity judgment. One important question remains probe in the first probe pair (for the distribution of raw regarding the computational mechanism underlying the response offsets, see the supplementary material available similarity-induced memory bias. One possibility is that at https://osf.io/79kbm/). Thus, a positive value indicates a WM is integrated with a probe if the probe is subjec- a response offset toward the first similar probe. For the tively similar to the WM (Fig. 5a; for a similar concep- baseline trials, the sign was randomly assigned. The mag- tualization, see Dubé et al., 2014). This integration can nitude of the memory bias was quantified as the mean of be accomplished via the formation of a joint density of the signed response offsets for each condition. We focused the two, resulting in a biased WM representation toward on trials with high-confidence memory reports (> 58% of a similar probe (Bae et al., 2015). Alternatively, the two trials, or > 25 trials, in all conditions; for the same pattern representations may be independently represented in of results with all trials included, see the supplementary WM, but participants may occasionally report the probe material available at https://osf.io/79kbm/). item instead of the memory item by mistake, especially when the probe is similar to the memory item (Fig. 5b). Importantly, this mixture density can also produce a Results shifted response distribution depending on the fre- Accuracy in the perceptual-comparison conditions was quency of the mistake (see the supplementary material near ceiling (shape task: M = 0.94, SD = 0.05 for the available at https://osf.io/79kbm/). We compared these same-side probe, M = 0.89, SD = 0.07 for the opposite- competing models by fitting them to the data obtained side probe; color task: M = 0.95, SD = 0.04 for the in Experiment 1. same-side probe, M = 0.92, SD = 0.04 for the opposite- side probe), thus confirming that subjective similarity Method of the probe was successfully controlled. As can be seen from Figures 4a and 4c, the same-side-probe con- Detailed descriptions of each model can be found in dition and the opposite-side-probe condition exhibited the supplementary material (https://osf.io/79kbm/). differential signed response offsets. The mean signed Here, we provide a summary. response offset for the same-side-probe condition For both the joint-density model and the mixture- exhibited a significant positive signed offset—shape density model, we assumed that both the memory (X ) task: M = 5.64°, t(27) = 7.17, p < .001, 95% CI = [4.02°, and the probe (X ) representations follow von Mises P Psychological Science 33(5) 823 a b Short Delay 20 0.03 Long Delay Same-Side 0.02 Opposite-Side 0.01 −10 0 90 180 −180 −90 Signed HC Response Offset (°) cd Short Delay 20 0.03 Long Delay Same-Side 0.02 Opposite-Side 0.01 −10 −180 −90 0 90 180 Signed HC Response Offset (°) Fig. 4. Results of the shape task (top row) and color task (bottom row) in Experiment 2. Probability distributions of signed response offsets for high-confidence (HC) reports (a, c) are shown for each of the four conditions in the shape task. For demonstration purposes, the response proportion for a given signed response offset was computed by calculating the mean response proportion across a 30° window centered around the signed response offset. Positive offsets indicate memory bias toward the first similar probe. The broken lines indicate within-subjects standard errors of the mean. Bias magnitude for HC memory reports in the shape task (b, d) is shown for same-side (pink) and opposite-side (blue) probe judgments. In each violin plot, the thick horizontal line indicates the mean across participants, and the width of the violin indicates the density of the data. Positive values indicate memory bias toward the first similar probe. distributions (denoted by φ) centered at the stimulus and the probe distributions were set by the actual stimulus value (S , S ) with some precision (κ , κ ): values (S , S ). The precision parameter for the memory M P M P M P item (κ ) was obtained by fitting a standard WM model pX || SX = φκ S , (1) (Zhang & Luck, 2008) to response offsets in the baseline () () MM MM M condition in Experiment 2. However, we fitted the preci- pX || SX = φκ S , (2) sion for the probe item within the model (κ ). Thus, the () () PP PP P P joint-density model has only one free parameter. When the probe item (Equation 2) is perceived to We fitted the joint-density model to each trial and be similar to the memory item (Equation 1), the joint- each participant separately for Experiments 1a and 1b density ( JD) model integrates the two representations data sets. We used only high-confidence memory in the following manner: reports to avoid contamination by guessing or lapses of attention. On a given trial for a given participant, we pX () || Sp() XS MM PP (3) constructed a joint-density distribution using S and S pX |, S S = () M P JD MP pX () || Sp() XS ∑ MM PP for that trial and κ for the participant and fitted the model by finding a probe precision (κ ) that minimized The joint-density model (Equation 3) has four param- the difference between the average human response eters. The two parameters for the center of the memory error collapsed across all the trials (including all the Color Shape Response Proportion Response Proportion HC Bias (°) HC Bias (°) 824 Fukuda et al. Memory Item Probe Item Joint Density p (x s , κ ) p (x s , κ ) p (x s , κ )p (x s , κ )     M M M P P P M M M P P P XX X Mixture Density αp (x s , κ ) Memory Item Probe Item  M M M + (1 − α)p (x s , κ ) p (x s , κ ) p (x s , κ )   P P P M M M P P P XX X Fig. 5. Two competing models of similarity-induced memory bias. In the joint-density model (a), noisy representation of a working memory item (X ) is assumed to follow a von Mises distribution centered at the sample stimulus (S , dashed black M M vertical line) with some precision (κ , left). Noisy representation of a probe item is also assumed to follow a von Mises distribution centered at the probe stimulus (S , dashed blue vertical line) with some precision (κ , center). When observers P P decide that the memory item and the probe item are perceptually similar, then the two items are integrated to produce a joint distribution (shown in red, right). As a result, the mean of the joint distribution is biased toward the probe item, as indicated by the dashed red vertical line. In the mixture-density model (b), the initial memory representation and the probe representation are assumed to follow von Mises distributions as in the joint-density model. However, this model assumes that some proportion of the memory reports is based on the probe representation. This can be accounted for by the mixture parameter (a). The original memory representation in this model is not biased, but the mean of the mixture distribution can be shifted toward the probe item (dashed red vertical line). Note that the schematics for both models depict the correspond- ing representational consequences for one trial with a given probe distance. For a simulation of multiple trials with varying probe distances as tested in the actual experiments, see the supplementary material available at https://osf.io/79kbm/. probe distances) and the average simulated response Results errors across all the simulated responses. Figure 6a shows simulated response-offset distributions The mixture-density model (Equation 4) combines the from the joint-density model and the mixture-density two distributions via a mixture parameter (a). Namely, model along with observed human data (Experiments 1a this model assumes that the final memory reports are and 1b). The peak of the simulated response distribution either memory-based or probe-based. The proportion of from the joint-density model was shifted positively, as in each is determined by the mixture parameter: the human data. However, the distributions from the mixture-density model were positively skewed without pX |, SS = αα pX || Sp +− 1 XS () () () () (4) MixM PM MP P this shift. This result suggests that the observed biases in the human data were more likely to be driven by All the other aspects of this model were identical to representational integration than probabilistic confusion. the joint-density model except that this model has an Formal model comparisons using measurements of the additional free parameter (a). Alpha was set to vary sum of log-likelihood, Akaike information criterion, and between 0 (0% memory-based reports) and 1 (100% Bayesian information criterion unanimously indicated memory-based reports). Probability Probability 0 a b Human Shape Working Memory Joint-Density Model Mixture-Density Model 2.0 2.0 2.0 1.5 1.5 1.5 1.0 1.0 1.0 0.5 0.5 0.5 0.0 0.0 0.0 05 10 15 20 −180 0 180 −180 0 180 −180 0 180 Mean Human Response Error (°) Signed Response Offset (°) Signed Response Offset (°) Signed Response Offset (°) Color Working Memory 2.0 2.0 2.0 1.5 1.5 1.5 1.0 1.0 1.0 0.5 0.5 0.5 0.0 0.0 0.0 05 10 15 20 −180 0 180 −180 0 180 −180 0 180 Signed Response Offset (°) Signed Response Offset (°) Signed Response Offset (°) Mean Human Response Error (°) Fig. 6. Observed and simulated memory bias in Experiments 1a (top row) and 1b (bottom row). Probability distributions of signed response errors (a) are shown separately for responses of human participants, predictions of the best-fitting joint-density model, and predictions of the best-fitting mixture-density model for each experiment. Correla- tions of bias magnitudes (b) between the simulated responses from the joint-density model and the observed responses from human participants in the shape and color tasks are shown for each of the 18 probe distances in each experiment. Positive values indicate a bias toward the probe item. The size of the dots represents the discretized physi- cal distance of the probe (smaller dots = closer probes, larger dots = further probes). Vertical error bars indicate ±1 bootstrapped standard error of the simulated responses. Horizontal error bars indicate ±1 bootstrapped standard error of human data. Experiment 1b Experiment 1a (Color Memory) (Shape Memory) Probability Probability Mean Joint-Density Model Mean Joint-Density Model Response Error (°) Response Error (°) 826 Fukuda et al. Table 1. Model-Fit Comparisons for Experiments 1a and 1b Experiment and model ∑ log-likelihood AIC BIC Experiment 1a: shape Joint density (κ ) –179.7096 391.4191 482.6021 Mixture density (a, κ ) –294.6754 653.3508 835.7167 Experiment 1b: color Joint density (κ ) –301.0448 634.0897 723.0092 Mixture density (a, κ ) –476.1254 1,016.251 1,194.09 Note: Free parameters are given in parentheses (κ = precision of probe representation; a = mixture parameter). Boldface indicates the preferred model. AIC = Akaike information criterion; BIC = Bayesian information criterion. that the joint-density model was preferred (Table 1). To of the same exclusion criteria used in Experiment 2. provide additional support for the joint-density model, we computed the correlation between the simulated and Procedure. The experiment was identical to Experi- observed bias magnitudes as a function of the 18 dis- ment 2 except for the following. There was one experi- cretized probe distances. First, we found that both the mental condition with one intervening similarity judgment observed and simulated bias magnitudes increased as the and one delay-matched baseline condition. The stimulus physical probe distance increased (observed: r = .66, p < onset asynchrony between the memory item and the .002 for color; r = .55, p = .018 for shape; simulated: r = response wheel was set to 4,000 ms for both conditions. .94, p < .001 for color; r = .94, p < .001 for shape). More Participants performed four blocks consisting of 15 trials importantly, the simulated bias magnitudes predicted the each of the baseline and experimental conditions (30 tri- observed bias magnitudes for both stimuli, even though als per block) in a pseudorandomized order. the model was not fitted separately for each distance (r = .75, p < .001 for color; r = .45, p = .037 for shape; Analysis. Memory precision was estimated by fitting Fig. 6b). the concentration parameter (κ) of a von Mises distribu- tion to the response offsets for high-confidence reports in the baseline condition (> 69% of trials, or > 41 trials) to Experiment 3 eliminate the effect of guessing. For the bias estimation, We found convincing evidence that representational we focused on trials with high-confidence response off- integration likely underlies the similarity-induced mem- sets (> 62% of trials, or > 37 trials; see the supplementary ory bias. One novel prediction of the joint-density material available at https://osf.io/79kbm/) for which model is that individuals with lower WM precision participants successfully identified the similar probe. The should exhibit a larger similarity-induced memory bias precision estimates were then correlated with the bias because of greater representational overlap between estimates. WM and probe representations (Fig. 5a; for a simula- tion, see the supplementary material available at https:// Results osf.io/79kbm/). To test this, we examined the correla- tion between individuals’ WM precision and the mag- Participants’ accuracy in the perceptual-comparison nitude of the similarity-induced memory bias using a task was near ceiling (proportion of correct responses: variant of the paradigm employed in Experiment 1. M = .96, SD = .03 for the shape task; M = .97, SD = .03 for the color task), and we replicated the similarity- induced memory bias (see the supplementary material Method available at https://osf.io/79kbm/). More importantly, Participants. A power analysis based on the effect size participants with high visual WM precision exhibited a obtained in Experiment 2 (r < –.32 between WM preci- smaller memory bias than those with low precision, sion and the bias magnitude) determined that at least 99 r(107) = −.37, p < .001 for color; r(105) = −.31, p = .001 samples would be necessary to detect such an effect with for shape (Fig. 7), as predicted by the joint-density a statistical power of .9. The informed-consent procedure model. This pattern was not changed when analyses was the same as in previous experiments, and 124 indi- were conducted with outliers that were initially excluded, viduals (97 female; mean age = 18.6 years) participated. r(108) = −.37, p < .001 for color; r(108) = −.33, p < .001 The data from 14 participants were removed on the basis for shape. Psychological Science 33(5) 827 Color Task Shape Task 10 10 0 0 −10 −10 0306090 120 0306090 120 HC Memory Precision (κ ) HC Memory Precision (κ ) M M Fig. 7. Results of the color task and shape task in Experiment 3. Each scatterplot depicts the relationship between bias for high-confidence (HC) reports and HC memory precision. Circles indicate individual participant data. Crosses indicate participant data excluded from the correlation as outliers. The solid lines indicate the best-fitting regressions. in different experimental blocks. We found that identi- General Discussion cal intervening inputs induced a larger memory bias in The present study revealed that the use of WM in per- compare blocks than in ignore blocks, and more impor- ceptual comparisons results in a systematic memory bias, tantly, we found that WM precision for the baseline thus demonstrating that WM is susceptible to interfer- condition (i.e., no intervening input) in both block ence even without task-irrelevant disruptions of active types was statistically indistinguishable. These results maintenance. This memory bias was particularly large demonstrated that the variability in WM precision alone when a visual input was subjectively similar to the WM. could not explain the memory bias and, thus, impli- Model comparisons showed that this similarity-induced cated a causal role of similarity judgments in modulat- memory bias was better characterized as representational ing the similarity-induced memory bias. integration between the WM and a subjectively similar The temporal stability of the similarity-induced mem- input (i.e., joint-density model) than probabilistic confu- ory bias also merits further discussion. Wang and col- sion between the two (i.e., mixture-density model). We leagues (2018) demonstrated that a WM representation also provided empirical support of this representational- can be distorted by similar perceptual inputs only when integration account by demonstrating that individuals they appear before attention returns to the WM from a with lower WM precision exhibited a larger memory bias secondary task. Although we did not use any secondary than those with higher precision. task to distract attention away from the original WM One may argue that the similarity-induced memory representation prior to the onset of the similar probe, bias is solely determined by WM precision, thus ques- it is possible that the similarity-induced memory bias tioning the causal role of perceptual comparisons in happens only when a WM is compared with a similar modulating the bias. More precisely, low-precision WM input too soon after encoding. We tested this possibility may be susceptible to greater stimulus-driven interfer- in a follow-up study (Saito et  al., 2021) by directly ence by a perceptual input, which may also result in manipulating the interval between memory encoding perceived similarity between them. In other words, per- and similarity judgments. Here, we replicated the ceived similarity between WM and a perceptual input similarity-induced memory bias even when participants may be a mere by-product of low WM precision. The performed a similarity judgment between a perceptual present study alone does not provide a strong test for input and a memory encoded 24 hr before by retrieving this possibility because we did not measure WM preci- it into WM (Atkinson & Shiffrin, 1968; Cowan, 1999; sion prior to the perceptual comparison. However, we Fukuda & Woodman, 2017). Furthermore, when par- tested this possibility in a follow-up study (Saito et al., ticipants retrieved a memory again 24 hr after a similar- 2020) by having participants either ignore or compare ity judgment, the memory reports remained biased. This intervening perceptual input with a WM representation temporal stability of the similarity-induced memory bias HC Bias (°) HC Bias (°) 828 Fukuda et al. This article has received the badge for Open Data. More is in stark contrast to the memory bias caused by a information about the Open Practices badges can be found temporary disruption of active WM maintenance, thus at http://www.psychologicalscience.org/publications/ suggesting that it is the direct usage of WM in percep- badges. tual comparisons that “locks in” the bias. Future studies will be necessary to determine the neural mechanisms that underlie these two distinct forms of memory biases. Our representational-integration account of WM bias Acknowledgments might offer a unifying explanation for other WM biases reported in the literature. For instance, past visual expe- We thank Steven J. Luck for the constructive comments on riences can bias subsequently encoded WM—a phe- the manuscript for this article. nomenon known as serial dependence in visual perception (Bae & Luck, 2019, 2020; Cicchini et  al., References 2018; Czoschke et al., 2019; C. Fischer et al., 2020; J. Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: Fischer & Whitney, 2014; Kiyonaga et  al., 2017). A proposed system and its control processes. In K. W. Although our finding that WM is biased by subsequently Spence & J. T. Spence (Eds.), The psychology of learning perceived inputs is different from serial dependence in and motivation: Advances in research and theory (Vol. 2, the direction of causality, serial dependence might also pp. 89–195). Academic Press. be explained by the representational integration trig- Bae, G.-Y. (2021). Neural evidence for categorical biases in location and orientation representations in a working gered by the subjective similarity between lingering memory task. NeuroImage, 240, Article 118366. https:// WM representations of past visual experience and cur- doi.org/10.1016/j.neuroimage.2021.118366 rent WM representations. Bae, G.-Y., & Luck, S. J. (2018). Dissociable decoding of Lastly, future studies should examine at what stage spatial attention and working memory from EEG the similarity-induced memory bias occurs. One pos- oscillations and sustained potentials. The Journal of sibility is that the memory bias might be introduced to Neuroscience, 38(2), 409–422. https://doi.org/10.1523/ WM as soon as its similarity to a novel input is per- JNEUROSCI.2860-17.2017 ceived. Alternatively, the WM might stay intact at the Bae, G.-Y., & Luck, S. J. (2019). Reactivation of previous time of the perceptual comparison, and the bias might experiences in a working memory task. Psycho logical be introduced at the time of memory report. To tease Science, 30(4), 587–595. https://doi.org/10.1177/095679 apart these two hypotheses, future studies should incor- 7619830398 Bae, G.-Y., & Luck, S. J. (2020). Serial dependence in vision: porate neural-decoding approaches (Bae, 2021; Bae & Merely encoding the previous-trial target is not enough. Luck, 2018; Rademaker et al., 2019) to track the content Psychonomic Bulletin & Review, 27(2), 293–300. https:// of WM as it is maintained, compared with a perceptual doi.org/10.3758/s13423-019-01678-7 input, and eventually reported. Bae, G.-Y., Olkkonen, M., Allred, S. R., & Flombaum, J. I. (2015). Why some colors appear more memorable than Transparency others: A model combining categories and particulars Action Editor: Barbara Knowlton in color working memory. Journal of Experimental Editor: Patricia J. Bauer Psychology: General, 144(4), 744–763. https://doi.org/10 Author Contributions .1037/xge0000076 K. Fukuda, A. E. Pereira, and H. Tsubomi designed the Bays, P. M., Catalao, R. F., & Husain, M. (2009). The precision study. A. E. Pereira and J. M. Saito collected the data. K. of visual working memory is set by allocation of a shared Fukuda analyzed the data. G.-Y. Bae conducted the model- resource. Journal of Vision, 9(10), Article 7. https://doi.org/ ing analysis. All the authors wrote the manuscript and 10.1167/9.10.7 approved the final manuscript for submission. Bennett, P. J., & Cortese, F. (1996). Masking of spatial fre- Declaration of Conflicting Interests quency in visual memory depends on distal, not retinal, The author(s) declared that there were no conflicts of frequency. Vision Research, 36(2), 233–238. https://doi interest with respect to the authorship or the publication .org/10.1016/0042-6989(95)00085-e of this article. Brainard, D. H. (1997). The Psychophysics Toolbox. Spatial Funding Vision, 10(4), 433–436. This research was supported by the Natural Sciences and Cicchini, G. M., Mikellidou, K., & Burr, D. C. (2018). The Engineering Research Council and the Connaught New functional role of serial dependence. Proceedings of the Researcher Award. Royal Society B: Biological Sciences, 285(1890), Article Open Practices 20181722. https://doi.org/10.1098/rspb.2018.1722 All data have been made publicly available via OSF and Cowan, N. (1999). An embedded-processes model of working can be accessed at https://osf.io/5v7sg/. The design and memory. In A. Miyake & P. Shah (Eds.), Models of working analysis plans for the experiments were not preregistered. memory: Mechanisms of active maintenance and executive Psychological Science 33(5) 829 control (pp. 62–101). Cambridge University Press. https:// Makovski, T., Sussman, R., & Jiang, Y. V. (2008). Orienting doi.org/10.1017/CBO9781139174909.006 attention in visual working memory reduces interference Czoschke, S., Fischer, C., Beitner, J., Kaiser, J., & Bledowski, from memory probes. Journal of Experimental Psychology: C. (2019). Two types of serial dependence in visual work- Learning, Memory, and Cognition, 34(2), 369–380. https:// ing memory. British Journal of Psychology, 110(2), 256– doi.org/10.1037/0278-7393.34.2.369 267. https://doi.org/10.1111/bjop.12349 Nemes, V. A., Parry, N. R., Whitaker, D., & McKeefry, D. J. Dubé, C., Zhou, F., Kahana, M. J., & Sekuler, R. (2014). (2012). The retention and disruption of color information Similarity-based distortion of visual short-term memory in human short-term visual memory. Journal of Vision, is due to perceptual averaging. Vision Research, 96, 8–16. 12(1), Article 26. https://doi.org/10.1167/12.1.26 https://doi.org/10.1016/j.visres.2013.12.016 Rademaker, R. L., Bloem, I. M., De Weerd, P., & Sack, A. T. Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007). (2015). The impact of interference on short-term memory G*Power 3: A flexible statistical power analysis program for visual orientation. Journal of Experimental Psychology: for the social, behavioral, and biomedical sciences. Human Perception and Performance, 41(6), 1650–1665. Behavior Research Methods, 39(2), 175–191. https://doi.org/10.1037/xhp0000110 Fischer, C., Czoschke, S., Peters, B., Rahm, B., Kaiser, J., & Rademaker, R. L., Chunharas, C., & Serences, J. T. (2019). Bledowski, C. (2020). Context information supports serial Coexisting representations of sensory and mnemonic infor- dependence of multiple visual objects across memory mation in human visual cortex. Nature Neuroscience, 22(8), episodes. Nature Communications, 11(1), Article 1932. 1336–1344. https://doi.org/10.1038/s41593-019-0428-x https://doi.org/10.1038/s41467-020-15874-w Rademaker, R. L., Park, Y. E., Sack, A. T., & Tong, F. (2018). Evidence Fischer, J., & Whitney, D. (2014). Serial dependence in visual of gradual loss of precision for simple features and complex perception. Nature Neuroscience, 17(5), 738–743. https:// objects in visual working memory. Journal of Experimental doi.org/10.1038/nn.3689 Psychology: Human Perception and Performance, 44(6), Fukuda, K., & Woodman, G. F. (2017). Visual working mem- 925–940. https://doi.org/10.1037/xhp0000491 ory buffers information retrieved from visual long-term Saito, J. M., Duncan, K., & Fukuda, K. (2021). Comparing memory. Proceedings of the National Academy of Sciences, visual memories to novel visual input risks lasting memory USA, 114(20), 5306–5311. https://doi.org/10.1073/pnas distortion. PsyArXiv. https://doi.org/10.31234/osf.io/xr4su .1617874114 Saito, J. M., Kolisnyk, M., & Fukuda, K. (2020). Perceptual Kiyonaga, A., Scimeca, J. M., Bliss, D. P., & Whitney, D. comparisons modulate memory biases induced by over- (2017). Serial dependence across perception, attention, lapping visual input. PsyArXiv. https://doi.org/10.31234/ and memory. Trends in Cognitive Sciences, 21(7), 493– osf.io/dqng3 497. https://doi.org/10.1016/j.tics.2017.04.011 Sun, S. Z., Fidalgo, C., Barense, M. D., Lee, A. C. H., Cant, J. S., Li, A. Y., Liang, J. C., Lee, A. C. H., & Barense, M. D. (2020). The & Ferber, S. (2017). Erasing and blurring memories: The validated circular shape space: Quantifying the visual simi- differential impact of interference on separate aspects of larity of shape. Journal of Experimental Psychology: General, forgetting. Journal of Experimental Psychology: General, 149(5), 949–966. https://doi.org/10.1037/xge0000693 146(11), 1606–1630. https://doi.org/10.1037/xge0000359 Luck, S. J., & Vogel, E. K. (2013). Visual working memory Teng, C., & Kravitz, D. J. (2019). Visual working memory capacity: From psychophysics and neurobiology to indi- directly alters perception. Nature Human Behaviour, 3(8), vidual differences. Trends in Cognitive Sciences, 17(8), 827–836. https://doi.org/10.1038/s41562-019-0640-4 391–400. https://doi.org/10.1016/j.tics.2013.06.006 Wang, B., Theeuwes, J., & Olivers, C. N. L. (2018). When Ma, W. J., Husain, M., & Bays, P. M. (2014). Changing con- shorter delays lead to worse memories: Task disruption cepts of working memory. Nature Neuroscience, 17(3), makes visual working memory temporarily vulnerable 347–356. https://doi.org/10.1038/nn.3655 to test interference. Journal of Experimental Psychology: Magnussen, S., & Greenlee, M. W. (1992). Retention and Learning, Memory, and Cognition, 44(5), 722–733. https:// disruption of motion information in visual short-term doi.org/10.1037/xlm0000468 memory. Journal of Experimental Psychology: Learning, Zhang, W., & Luck, S. J. (2008). Discrete fixed-resolution Memory, and Cognition, 18(1), 151–156. https://doi.org/ representations in visual working memory. Nature, 10.1037//0278-7393.18.1.151 453(7192), 233–235. https://doi.org/10.1038/nature06860 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Psychological Science SAGE

Working Memory Content Is Distorted by Its Use in Perceptual Comparisons

Loading next page...
 
/lp/sage/working-memory-content-is-distorted-by-its-use-in-perceptual-7V8Ox5WtS0

References (35)

Publisher
SAGE
Copyright
© The Author(s) 2022
ISSN
0956-7976
eISSN
1467-9280
DOI
10.1177/09567976211055375
Publisher site
See Article on Publisher Site

Abstract

Visual information around us is rarely static. To perform a task in such a dynamic environment, we often have to compare current visual input with our working memory (WM) representation of the immediate past. However, little is known about what happens to a WM representation when it is compared with perceptual input. To test this, we asked young adults (N = 170 total in three experiments) to compare a new visual input with a WM representation prior to reporting the WM representation. We found that the perceptual comparison biased the WM report, especially when the input was subjectively similar to the WM representation. Furthermore, using computational modeling and individual- differences analyses, we found that this similarity-induced memory bias was driven by representational integration, rather than incidental confusion, between the WM representation and subjectively similar input. Together, our findings highlight a novel source of WM distortion and suggest a general mechanism that determines how WM interacts with new visual input. Keywords working memory, memory distortion, individual differences, open data Received 10/16/20; Revision accepted 9/22/21 Many daily activities rely on our ability to maintain during maintenance (e.g., Bennett & Cortese, 1996; accurate mental representations of the immediate past Magnussen & Greenlee, 1992; Nemes et  al., 2012; Sun to guide our behavior. For example, consider a Good et  al., 2017; Teng & Kravitz, 2019) and at the time of Samaritan who witnesses a phone falling out of a pass- memory reporting (Makovski et  al., 2008; Wang et  al., erby’s pocket on a busy street. The Good Samaritan 2018). For example, in a previous study, participants were would look down to pick up the phone in hopes of asked to maintain a color in WM while completing a rapid returning it to its owner. After doing so, they would serial visual presentation (RSVP) task in which a stream need to find the owner again by comparing their mem- of letters was presented on a differing color patch (Teng ory of the owner’s appearance with the other people & Kravitz, 2019). The researchers found that the color they see on the street. Past studies have demonstrated representation maintained in WM was distorted by the that working memory (WM) can actively maintain a task-irrelevant color patch, especially when the two colors small amount of task-relevant information (e.g., the were physically similar to each other. Therefore, WM owner’s appearance) over a brief delay when the infor- seems vulnerable to perceptual interference. mation is not perceptually present (e.g., while looking However, the true nature of this interference remains down on the street to pick up the phone; Luck & Vogel, unknown. For example, it is unclear whether the mem- 2013; Ma et al., 2014). However, little is known as to ory distortion occurs if a WM representation remains whether WM remains intact as it is used for memory- guided perceptual comparisons. Corresponding Author: Although one might assume that WM remains intact as Keisuke Fukuda, University of Toronto Mississauga, Department of it is used, past studies have shown that WM can be inter- Psychology fered with by subsequent perceptual inputs introduced Email: keisuke.fukuda@utoronto.ca Psychological Science 33(5) 817 within the focus of internal attention. Given that the intervening RSVP task used by Teng and Kravitz (2019) Statement of Relevance drew attention away from WM maintenance, the mem- Imagine a person who witnesses a hit-and-run ory distortion may be contingent on disruption of active traffic accident. When the witness tries to identify WM maintenance (Makovski et al., 2008; Wang et al., the license plate number of the car involved, a 2018). It is also possible that the WM representation bus occludes their sight of the car. When the bus was not necessarily distorted but was occasionally passes, they need to find the car by comparing replaced by the intervening stimulus instead (i.e., swap the cars they see now with the memory of the car errors), making WM reports appear biased (Bays et al., that committed the accident. In doing so, they 2009). Lastly, because previous studies did not measure might assume that their memory of the car remains the perceived similarity between WM and intervening intact as it is compared with the cars they see stimuli, it is not known whether the interference was now. However, the present study shows that this dictated by the physical similarity or perceived similar- assumption is simply not valid: As a memory rep- ity between them. resentation of the immediate past (i.e., working We addressed these questions in the present study memory) is compared with new perceptual infor- by having participants remember a single visual feature mation, the memory report becomes systemati- (i.e., color or shape) in WM and compare it with new cally biased toward perceptually similar input. stimuli. Therefore, the WM had to be actively main- Thus, our findings reveal an overlooked vulner- tained so that it could be compared with new stimuli. ability in working memory and suggest that the In three experiments, we found that WM reports were witness may misreport the license plate of a simi- distorted, especially when participants perceived the lar, but different, car to the police. new stimuli as similar to the WM. Critically, we found that this similarity-induced memory bias was larger fol- lowing perceived similarity than perceived dissimilarity reported large effect sizes (Cohen’s d > 1.5). Because we of new stimuli, even after controlling for their physical used different stimuli, we anticipated a medium to large similarity. We further validated the systematic effect of effect size (Cohen’s d = 0.8). A power analysis (conducted subjective similarity by demonstrating that memory in G*Power Version 3.1; Faul et al., 2007) determined that biases induced by an initial similarity judgment were at least 15 samples would be necessary to detect such an “canceled out” by an additional judgment that biased effect with a statistical power of .8. We recruited 16 par- the WM in an opposing direction. Computational mod- ticipants (11 female; mean age = 23.2 years) for Experi- eling revealed that the similarity-induced memory bias ment 1a and 16 participants (10 female; mean age = 22.6 was better accounted for by representational integration years) for Experiment 1b. Participants provided informed between WM and perceptually similar input than by consent, and the study was approved by the University of probabilistic swap errors. Thus, our findings reveal an Toronto Research Ethics Board. overlooked window of vulnerability in WM and suggest a general mechanism that determines how a new per- Apparatus, stimuli, and procedure. Stimuli were gen- ceptual input interacts with WM (e.g., Kiyonaga et al., erated in MATLAB (The MathWorks, Natick, MA) using the 2017; Rademaker et al., 2019; Sun et al., 2017). Psychophysics Toolbox (Brainard, 1997) and were pre- sented at 60 Hz on an LCD monitor. Viewing distance was Experiments 1a and 1b approximately 60 cm. The shape stimulus set used in Experiment 1a contained 360 shapes whose circular conti- To examine the consequences of comparing WM with nuity was empirically validated (Li et al., 2020). The color perceptual inputs, we had participants remember a stimuli used in Experiment 1b comprised 360 equally simple stimulus (i.e., shape in Experiment 1a and color spaced color values sampled from Commission Internatio- in Experiment 1b) and compare it with a new percep- nale de l’Éclairage (CIE) L*a*b* space with a* centered at 20 tual input (Fig. 1). Subsequently, participants reported and b* centered at 38 with a radius of 60. L* was set to 70. their memory of the original stimulus as precisely as Participants completed multiple trials in the baseline possible. Here, we predicted that the WM report would and perceptual-comparison conditions (Fig. 1b). In the be biased toward the new input, especially when the baseline condition, participants were presented with new input was subjectively similar to the WM. one stimulus—either a shape (3.8° × 3.8°) or a colored circle (5.2° diameter) for 800 ms. After a 3,200-ms reten- Method tion interval, participants were presented with a circular stimulus wheel (15.4° diameter). For the shape task, Participants. Related experiments examining the mem- the stimulus wheel was composed of 18 equidistant ory bias for orientation memory (Rademaker et al., 2015) 818 Fukuda et al. Shape Space Color Space “Remember” “Recall” “Adjust” “Confidence?” 1 2 3 “Similar or Not?” “Remember” “Recall” “Adjust” “Confidence?” 1 2 3 800 ms 800 ms 1,600 ms Until Response Until Response Until Response + 800-ms Delay Fig. 1. Stimulus spaces and example trial procedures from Experiments 1a and 1b. The stimulus spaces used for Experiments 1a (shape task) and 1b (color task) are shown in (a). Trial procedures from the shape task (b) are shown separately for the baseline and perceptual- comparison conditions. In each condition, participants were first presented with a target shape that they attempted to remember across a brief maintenance interval. Following the maintenance interval, participants reported the target shape by rotating a line indicator to select from a continuous wheel. Participants then adjusted their selection, if desired, to best match their memory before submitting their report with a confidence rating. In the perceptual-comparison condition, participants also performed a perceptual comparison on a probe shape that was presented during the maintenance interval by indicating whether the probe shape was similar or dissimilar to the target shape. The color task was identical to the shape task except for the stimulus type and the stimulus wheel. shapes (20° apart in the shape space). For the color looked similar or dissimilar to the original stimulus by task, it was a color wheel composed of the 360 colors pressing a key. The probe stayed on the screen for 1,600 (1° per color). To report the remembered stimulus, they ms. The probe was randomly sampled from 16° to 105° first selected an item that best matched the original away from the original stimulus on each trial. After an stimulus from the stimulus wheel by rotating a line 800-ms delay following the probe presentation, partici- indicator using the left and right arrow keys on the pants reported the original memory item as in the base- keyboard. After the indicator pointed toward the desired line condition. Participants performed 12 experimental item, participants pressed the space bar to confirm the blocks, and each block contained 12 baseline and 36 selection. The selected item then appeared at the center perceptual-comparison trials. of the screen for optional refinements using the left and right arrow keys. When satisfied, participants indicated Analysis. For each trial, a signed response offset was their confidence in their report (1 = high confidence, calculated in relation to the probe. A positive sign indi- 2 = low confidence, 3 = no confidence) by pressing cated a response offset toward the probe (for the distri- one of three keys on the keyboard. No time limit was bution of raw response offsets, see the supplementary imposed on the memory report in order to emphasize material available at https://osf.io/79kbm/). For the base- accuracy. line trials, the sign of the response offset was randomly In the perceptual-comparison condition, a probe assigned. To quantify the magnitude of the bias, we com- stimulus was presented 800 ms after the offset of the puted the mean of the signed response offsets for each original stimulus, and participants indicated whether it condition. We focused on trials with high-confidence Perceptual Baseline Comparison Psychological Science 33(5) 819 memory reports (> 68% of trials, or > 83 trials, in all con- modified paradigm. More precisely, participants clicked ditions; for the same pattern of results with all trials a mouse to report WM and performed two-alternative included, see the supplementary material available at forced-choice judgments for perceptual comparisons https://osf.io/79kbm/). (Fig. 3). More importantly, considering that many To isolate the effect of perceived similarity on the behaviors rely on the maintenance of unbiased WM memory bias, we first identified the probe distances that representations, we examined whether the similarity- resulted in both similar and dissimilar judgments on induced memory bias can be corrected by biasing WM separate trials within subjects (ambivalent probe dis- in the opposite direction in a subsequent perceptual tances). For each ambivalent probe distance, we calcu- judgment. If the bias can be corrected by subsequent lated the mean bias magnitude following similar and perceptual judgments, then its magnitude should be dissimilar judgments separately. The mean bias magni- smaller when the similar probes in two consecutive tudes following similar and dissimilar judgments were judgments were sampled from opposing directions (i.e., then averaged across all ambivalent probe distances. This clockwise and counterclockwise) than from the same procedure isolated the effect of perceived probe similar- direction. ity (indicated by participants’ judgments) on the bias magnitude while equating the effect of physical probe Method similarity (determined by the sampling procedure). Participants. On the basis of the results of Experiment 1 (Cohen’s d > 0.73), we anticipated a medium to large Results effect size (Cohen’s d = 0.73). A power analysis (con- Both shape (Experiment 1a) and color (Experiment 1b) ducted in G*Power Version 3.1; Faul et  al., 2007) deter- memory reports exhibited an attraction bias (> 0°) mined that 22 samples were necessary to detect such an toward the probe when the probe was judged to be effect with a statistical power of .9. After providing written similar to the memory item—shape: M = 7.61°, t(15) = informed consent to protocols approved by the University 4.70, p < .001, 95% confidence interval (CI) = [4.28°, of Toronto Research Ethics Board, 32 individuals (24 10.94°], Cohen’s d = 1.18 (Figs. 2a and 2b); color: M = female; mean age = 19.3 years) participated. The data from 8.14°, t(15) = 5.47, p < .001, 95% CI = [4.97°, 11.32°], four participants were removed because they did not com- Cohen’s d = 1.37 (Figs. 2e and 2f ). We also found some plete the experiment (n = 1), failed to follow instructions evidence for an attraction bias when the probe was (n = 1), or too infrequently made high-confidence mem- judged to be dissimilar to the original memory item— ory reports (i.e., < 15% for the baseline conditions; n = 2). shape: M = 1.49°, t(15) = 1.48, p = .159, 95% CI = [–0.65°, As a result, 28 participants’ data were analyzed. 3.64°], Cohen’s d = 0.37; color: M = 3.03°, t(15) = 2.93, p = .01, 95% CI = [0.83°, 5.12°], Cohen’s d = 0.73—but Procedure. The procedure was similar to that of Experi- the bias magnitude was significantly smaller than when ment 1. Participants performed four blocks of the color the probe was judged to be similar—similar shape ver- task and four blocks of the shape task in a pseudoran- sus dissimilar shape: M = 5.42°, t(15) = 3.14, p = .007, domized order. In the baseline conditions (Fig. 3a), par- 95% CI = [1.74°, 9.10°], Cohen’s d = 0.78; similar color ticipants remembered the original stimulus over a 500-ms versus dissimilar color: M = 5.11°, t(15) = 2.98, p = .009, (short delay) or a 5,500-ms (long delay) interval. Two 95% CI = [1.45°, 8.77°], Cohen’s d = 0.74. The trials with delays were introduced to establish the baseline response- ambivalent probe distances (shape task: mean range of offset distributions for the perceptual-comparison condi- probe distance = 20.1°–97.7°; color task: mean range of tions (for an example of the effect of delay on WM probe distance = 19.3°–97.6°) produced larger attraction precision, see Rademaker et  al., 2018). After the delay, biases when the probes were perceived to be similar participants used a mouse to click on an item that best rather than dissimilar to the original memory items— matched the original stimulus, fine-tuned their response shape task: ΔM = 6.19°, t(15) = 3.01, p = .009, 95% CI = using key presses, and then reported their confidence as [1.81°, 10.58°], Cohen’s d = 0.75 (Figs. 2c and 2d); color in Experiment 1. task: ΔM = 10.92°, t(15) = 4.50, p < .001, 95% CI = [5.75°, In the perceptual-comparison conditions, partici- 16.08°], Cohen’s d = 1.13 (Figs. 2g and 2h). pants performed two intervening perceptual compari- sons 500 ms after the original stimulus offset. Because subjective similarity of a given perceptual probe varied Experiment 2 across trials, we controlled the subjective similarity of Experiment 1 demonstrated that WM reports were a physically similar probe (e.g., 30° away from the biased when WM was directly compared with a subjec- target) by presenting it together with a physically dis- tively similar probe. In Experiment 2, we sought to similar probe (180° away from the similar probe) and replicate this similarity-induced memory bias in a had participants identify the more similar probe from HC Bias (°) for HC Bias (°) for Ambivalent Probe Trials Ambivalent Probe Trials ab cd 0.03 0.03 40 40 Baseline Similar 0.02 0.02 Dissimilar 20 20 0.01 0.01 0 0 −20 −20 −180 −90 0 90 180 −180 −90 0 90 180 Signed Response Offset for HC Reports (°) Signed Response Offset for HC Reports (°) ef gh 0.03 0.03 40 40 Baseline Similar 0.02 0.02 Dissimilar 20 20 0.01 0.01 0 0 0 0 −20 −20 0 90 180 0 90 180 −180 −90 −180 −90 Signed Response Offset for HC Reports (°) Signed Response Offset for HC Reports (°) Fig. 2. Results of Experiments 1a (top row) and 1b (bottom row). Probability distributions of signed response offsets for high-confidence (HC) reports (a, e) are shown for the baseline condition and for similar and dissimilar probe judgments in the perceptual-comparison condition. For demonstration purposes, the response proportion for a given signed response offset was computed by calculating the mean response proportion across a 30° window centered around the signed response offset. Positive offsets indicate memory bias toward the first similar probe. The broken lines indicate within-subjects standard errors of the mean. Bias magnitude for HC memory reports (b, f) is shown for similar (pink) and dissimilar (blue) probe judgments. In each violin plot, the thick horizontal line indicates the mean across participants, and the width of the violin indicates the density of the data. Positive values indicate memory bias toward the probe. Probability distributions of signed response offsets for HC reports for ambivalent probe trials (c, g) are shown for the baseline condition and for similar and dissimilar probe judgments in the perceptual-comparison condition. Bias magnitude for HC memory reports for ambivalent probe trials (d, h) is shown for similar (pink) and dissimilar (blue) probe judgments. In each violin plot, the thick horizontal line indicates the mean across participants, and the width of the violin indicates the density of the data. Positive values indicate memory bias toward the probe. Experiment 1b (Color) Experiment 1a (Shape) Response Proportion Response Proportion HC Bias (°) HC Bias (°) Response Proportion for Response Proportion for Ambivalent Probe Trials Ambivalent Probe Trials 821 “Remember” “Recall” “Adjust” “Confidence?” Confidence Report 1: High Confidence Short Delay 2: Low Confidence 3: No Confidence 1 2 3 Until Response Until Response Until Response “Remember” “Recall” “Adjust” “Confidence?” Long Delay 1 2 3 “Remember” “More Similar?” “More Similar?” “Recall” “Adjust” “Confidence?” Same-Side Probe 1 2 3 “Remember” “More Similar?” “Recall” “Adjust” “More Similar?” “Confidence?” Opposite-Side Probe 1 2 3 1,500 ms 500 ms 2,000 ms 2,000 ms Until Response Until Response Until Response + 500-ms Delay + 500-ms Delay Same-Side Probe Opposite-Side Probe Memory Item Memory Item 1st Probe Pair 2nd Probe Pair Fig. 3. Trial procedures and probe-sampling procedures for Experiment 2. Trial procedures from the color task (a) are shown separately for the baseline (short delay, long delay) and experimental (same-side probe, opposite-side probe) conditions. In each condition, participants were first presented with a target color that they attempted to remember across a brief maintenance interval. Following the maintenance interval, participants reported the target color by clicking on a continuous wheel. Participants then adjusted their selection, if desired, to best match their memory before submitting their report with a confidence rating. In the experimental conditions, participants also performed two consecutive perceptual comparisons during the maintenance interval. In each comparison, participants selected which of two simultaneously presented probe colors was more similar to the target color. The sampling of the similar probes is depicted in (b). In the same-side-probe condition, the similar probes were both sampled from the same side of the circular color space relative to the target. In the opposite-side-probe condition, the similar probes were sampled from opposite sides of the target. The dissimilar probes were sampled 180° from the similar probes in both pairs. The shape task was identical to the color task except for the stimulus type and the stimulus wheel. 822 Fukuda et al. the pair (Fig. 3). In each comparison, two probes were 7.25°], Cohen’s d = 1.36 (Fig. 4b); color task: M = 6.11°, presented on each side of the screen (5.2° from the t(27) = 6.30, p < .001, 95% CI = [4.12°, 8.09°], Cohen’s center of the screen), and they reported which probe d = 1.19 (Fig. 4d)—indicating that the memory reports looked more similar to the original stimulus by pressing were attracted toward the similar probes. In contrast, either the left or right arrow key. One of the probes the mean signed response offset for the opposite-side- was randomly sampled from ±16° to 45° away from the probe condition exhibited a nonsignificant negative original stimulus (i.e., similar probe). The other probe signed offset for shape, M = −0.36°, t(27) = −0.78, p = was sampled 180° away from the similar probe (i.e., .441, 95% CI = [–1.32°, 0.59°], Cohen’s d = −0.15, and dissimilar probe). The two probes remained on the a small but significant negative signed offset for color, screen for 2,000 ms regardless of the report. After the M = −1.29°, t(27) = 2.20, p = .037, 95% CI = [–2.49°, offset of the first pair of probes, a 500-ms delay fol- –0.09°], Cohen’s d = −0.42, indicating that the memory lowed, and the second pair of probes was presented reports were, if anything, biased away from the first for another similarity judgment (2,000 ms). After another similar probe. The magnitude of the bias (i.e., absolute 500-ms delay, participants reported the original stimulus values) for the same-side-probe condition was statisti- in the same manner as the baseline conditions. cally greater than the magnitude of the bias for the The two perceptual-comparison conditions differed opposite-side-probe conditions—shape task: ΔM = in how the similar probes were sampled (Fig. 3b). In 5.27°, t(27) = 7.46, p < .001, 95% CI = [3.82°, 6.72°], the same-side-probe condition, similar probes in each Cohen’s d = 1.41; color task: ΔM = 4.82°, t(27) = 3.88, pair were sampled from the same side of the stimulus p < .001, 95% CI = [2.27°, 7.37°], Cohen’s d = 0.73. space relative to the memory item. In the opposite-side- probe condition, similar probes in each pair were sam- Modeling of the Similarity-Induced pled from opposite sides of the stimulus space relative Memory Bias to the memory item. Participants completed 40 trials for each condition in a pseudorandomized order. Experiment 2 not only replicated the similarity-induced memory bias in a modified paradigm but also demon- Analysis. For each trial, a signed response offset was cal- strated that it could be corrected by an additional simi- culated as the response offset in the direction of the similar larity judgment. One important question remains probe in the first probe pair (for the distribution of raw regarding the computational mechanism underlying the response offsets, see the supplementary material available similarity-induced memory bias. One possibility is that at https://osf.io/79kbm/). Thus, a positive value indicates a WM is integrated with a probe if the probe is subjec- a response offset toward the first similar probe. For the tively similar to the WM (Fig. 5a; for a similar concep- baseline trials, the sign was randomly assigned. The mag- tualization, see Dubé et al., 2014). This integration can nitude of the memory bias was quantified as the mean of be accomplished via the formation of a joint density of the signed response offsets for each condition. We focused the two, resulting in a biased WM representation toward on trials with high-confidence memory reports (> 58% of a similar probe (Bae et al., 2015). Alternatively, the two trials, or > 25 trials, in all conditions; for the same pattern representations may be independently represented in of results with all trials included, see the supplementary WM, but participants may occasionally report the probe material available at https://osf.io/79kbm/). item instead of the memory item by mistake, especially when the probe is similar to the memory item (Fig. 5b). Importantly, this mixture density can also produce a Results shifted response distribution depending on the fre- Accuracy in the perceptual-comparison conditions was quency of the mistake (see the supplementary material near ceiling (shape task: M = 0.94, SD = 0.05 for the available at https://osf.io/79kbm/). We compared these same-side probe, M = 0.89, SD = 0.07 for the opposite- competing models by fitting them to the data obtained side probe; color task: M = 0.95, SD = 0.04 for the in Experiment 1. same-side probe, M = 0.92, SD = 0.04 for the opposite- side probe), thus confirming that subjective similarity Method of the probe was successfully controlled. As can be seen from Figures 4a and 4c, the same-side-probe con- Detailed descriptions of each model can be found in dition and the opposite-side-probe condition exhibited the supplementary material (https://osf.io/79kbm/). differential signed response offsets. The mean signed Here, we provide a summary. response offset for the same-side-probe condition For both the joint-density model and the mixture- exhibited a significant positive signed offset—shape density model, we assumed that both the memory (X ) task: M = 5.64°, t(27) = 7.17, p < .001, 95% CI = [4.02°, and the probe (X ) representations follow von Mises P Psychological Science 33(5) 823 a b Short Delay 20 0.03 Long Delay Same-Side 0.02 Opposite-Side 0.01 −10 0 90 180 −180 −90 Signed HC Response Offset (°) cd Short Delay 20 0.03 Long Delay Same-Side 0.02 Opposite-Side 0.01 −10 −180 −90 0 90 180 Signed HC Response Offset (°) Fig. 4. Results of the shape task (top row) and color task (bottom row) in Experiment 2. Probability distributions of signed response offsets for high-confidence (HC) reports (a, c) are shown for each of the four conditions in the shape task. For demonstration purposes, the response proportion for a given signed response offset was computed by calculating the mean response proportion across a 30° window centered around the signed response offset. Positive offsets indicate memory bias toward the first similar probe. The broken lines indicate within-subjects standard errors of the mean. Bias magnitude for HC memory reports in the shape task (b, d) is shown for same-side (pink) and opposite-side (blue) probe judgments. In each violin plot, the thick horizontal line indicates the mean across participants, and the width of the violin indicates the density of the data. Positive values indicate memory bias toward the first similar probe. distributions (denoted by φ) centered at the stimulus and the probe distributions were set by the actual stimulus value (S , S ) with some precision (κ , κ ): values (S , S ). The precision parameter for the memory M P M P M P item (κ ) was obtained by fitting a standard WM model pX || SX = φκ S , (1) (Zhang & Luck, 2008) to response offsets in the baseline () () MM MM M condition in Experiment 2. However, we fitted the preci- pX || SX = φκ S , (2) sion for the probe item within the model (κ ). Thus, the () () PP PP P P joint-density model has only one free parameter. When the probe item (Equation 2) is perceived to We fitted the joint-density model to each trial and be similar to the memory item (Equation 1), the joint- each participant separately for Experiments 1a and 1b density ( JD) model integrates the two representations data sets. We used only high-confidence memory in the following manner: reports to avoid contamination by guessing or lapses of attention. On a given trial for a given participant, we pX () || Sp() XS MM PP (3) constructed a joint-density distribution using S and S pX |, S S = () M P JD MP pX () || Sp() XS ∑ MM PP for that trial and κ for the participant and fitted the model by finding a probe precision (κ ) that minimized The joint-density model (Equation 3) has four param- the difference between the average human response eters. The two parameters for the center of the memory error collapsed across all the trials (including all the Color Shape Response Proportion Response Proportion HC Bias (°) HC Bias (°) 824 Fukuda et al. Memory Item Probe Item Joint Density p (x s , κ ) p (x s , κ ) p (x s , κ )p (x s , κ )     M M M P P P M M M P P P XX X Mixture Density αp (x s , κ ) Memory Item Probe Item  M M M + (1 − α)p (x s , κ ) p (x s , κ ) p (x s , κ )   P P P M M M P P P XX X Fig. 5. Two competing models of similarity-induced memory bias. In the joint-density model (a), noisy representation of a working memory item (X ) is assumed to follow a von Mises distribution centered at the sample stimulus (S , dashed black M M vertical line) with some precision (κ , left). Noisy representation of a probe item is also assumed to follow a von Mises distribution centered at the probe stimulus (S , dashed blue vertical line) with some precision (κ , center). When observers P P decide that the memory item and the probe item are perceptually similar, then the two items are integrated to produce a joint distribution (shown in red, right). As a result, the mean of the joint distribution is biased toward the probe item, as indicated by the dashed red vertical line. In the mixture-density model (b), the initial memory representation and the probe representation are assumed to follow von Mises distributions as in the joint-density model. However, this model assumes that some proportion of the memory reports is based on the probe representation. This can be accounted for by the mixture parameter (a). The original memory representation in this model is not biased, but the mean of the mixture distribution can be shifted toward the probe item (dashed red vertical line). Note that the schematics for both models depict the correspond- ing representational consequences for one trial with a given probe distance. For a simulation of multiple trials with varying probe distances as tested in the actual experiments, see the supplementary material available at https://osf.io/79kbm/. probe distances) and the average simulated response Results errors across all the simulated responses. Figure 6a shows simulated response-offset distributions The mixture-density model (Equation 4) combines the from the joint-density model and the mixture-density two distributions via a mixture parameter (a). Namely, model along with observed human data (Experiments 1a this model assumes that the final memory reports are and 1b). The peak of the simulated response distribution either memory-based or probe-based. The proportion of from the joint-density model was shifted positively, as in each is determined by the mixture parameter: the human data. However, the distributions from the mixture-density model were positively skewed without pX |, SS = αα pX || Sp +− 1 XS () () () () (4) MixM PM MP P this shift. This result suggests that the observed biases in the human data were more likely to be driven by All the other aspects of this model were identical to representational integration than probabilistic confusion. the joint-density model except that this model has an Formal model comparisons using measurements of the additional free parameter (a). Alpha was set to vary sum of log-likelihood, Akaike information criterion, and between 0 (0% memory-based reports) and 1 (100% Bayesian information criterion unanimously indicated memory-based reports). Probability Probability 0 a b Human Shape Working Memory Joint-Density Model Mixture-Density Model 2.0 2.0 2.0 1.5 1.5 1.5 1.0 1.0 1.0 0.5 0.5 0.5 0.0 0.0 0.0 05 10 15 20 −180 0 180 −180 0 180 −180 0 180 Mean Human Response Error (°) Signed Response Offset (°) Signed Response Offset (°) Signed Response Offset (°) Color Working Memory 2.0 2.0 2.0 1.5 1.5 1.5 1.0 1.0 1.0 0.5 0.5 0.5 0.0 0.0 0.0 05 10 15 20 −180 0 180 −180 0 180 −180 0 180 Signed Response Offset (°) Signed Response Offset (°) Signed Response Offset (°) Mean Human Response Error (°) Fig. 6. Observed and simulated memory bias in Experiments 1a (top row) and 1b (bottom row). Probability distributions of signed response errors (a) are shown separately for responses of human participants, predictions of the best-fitting joint-density model, and predictions of the best-fitting mixture-density model for each experiment. Correla- tions of bias magnitudes (b) between the simulated responses from the joint-density model and the observed responses from human participants in the shape and color tasks are shown for each of the 18 probe distances in each experiment. Positive values indicate a bias toward the probe item. The size of the dots represents the discretized physi- cal distance of the probe (smaller dots = closer probes, larger dots = further probes). Vertical error bars indicate ±1 bootstrapped standard error of the simulated responses. Horizontal error bars indicate ±1 bootstrapped standard error of human data. Experiment 1b Experiment 1a (Color Memory) (Shape Memory) Probability Probability Mean Joint-Density Model Mean Joint-Density Model Response Error (°) Response Error (°) 826 Fukuda et al. Table 1. Model-Fit Comparisons for Experiments 1a and 1b Experiment and model ∑ log-likelihood AIC BIC Experiment 1a: shape Joint density (κ ) –179.7096 391.4191 482.6021 Mixture density (a, κ ) –294.6754 653.3508 835.7167 Experiment 1b: color Joint density (κ ) –301.0448 634.0897 723.0092 Mixture density (a, κ ) –476.1254 1,016.251 1,194.09 Note: Free parameters are given in parentheses (κ = precision of probe representation; a = mixture parameter). Boldface indicates the preferred model. AIC = Akaike information criterion; BIC = Bayesian information criterion. that the joint-density model was preferred (Table 1). To of the same exclusion criteria used in Experiment 2. provide additional support for the joint-density model, we computed the correlation between the simulated and Procedure. The experiment was identical to Experi- observed bias magnitudes as a function of the 18 dis- ment 2 except for the following. There was one experi- cretized probe distances. First, we found that both the mental condition with one intervening similarity judgment observed and simulated bias magnitudes increased as the and one delay-matched baseline condition. The stimulus physical probe distance increased (observed: r = .66, p < onset asynchrony between the memory item and the .002 for color; r = .55, p = .018 for shape; simulated: r = response wheel was set to 4,000 ms for both conditions. .94, p < .001 for color; r = .94, p < .001 for shape). More Participants performed four blocks consisting of 15 trials importantly, the simulated bias magnitudes predicted the each of the baseline and experimental conditions (30 tri- observed bias magnitudes for both stimuli, even though als per block) in a pseudorandomized order. the model was not fitted separately for each distance (r = .75, p < .001 for color; r = .45, p = .037 for shape; Analysis. Memory precision was estimated by fitting Fig. 6b). the concentration parameter (κ) of a von Mises distribu- tion to the response offsets for high-confidence reports in the baseline condition (> 69% of trials, or > 41 trials) to Experiment 3 eliminate the effect of guessing. For the bias estimation, We found convincing evidence that representational we focused on trials with high-confidence response off- integration likely underlies the similarity-induced mem- sets (> 62% of trials, or > 37 trials; see the supplementary ory bias. One novel prediction of the joint-density material available at https://osf.io/79kbm/) for which model is that individuals with lower WM precision participants successfully identified the similar probe. The should exhibit a larger similarity-induced memory bias precision estimates were then correlated with the bias because of greater representational overlap between estimates. WM and probe representations (Fig. 5a; for a simula- tion, see the supplementary material available at https:// Results osf.io/79kbm/). To test this, we examined the correla- tion between individuals’ WM precision and the mag- Participants’ accuracy in the perceptual-comparison nitude of the similarity-induced memory bias using a task was near ceiling (proportion of correct responses: variant of the paradigm employed in Experiment 1. M = .96, SD = .03 for the shape task; M = .97, SD = .03 for the color task), and we replicated the similarity- induced memory bias (see the supplementary material Method available at https://osf.io/79kbm/). More importantly, Participants. A power analysis based on the effect size participants with high visual WM precision exhibited a obtained in Experiment 2 (r < –.32 between WM preci- smaller memory bias than those with low precision, sion and the bias magnitude) determined that at least 99 r(107) = −.37, p < .001 for color; r(105) = −.31, p = .001 samples would be necessary to detect such an effect with for shape (Fig. 7), as predicted by the joint-density a statistical power of .9. The informed-consent procedure model. This pattern was not changed when analyses was the same as in previous experiments, and 124 indi- were conducted with outliers that were initially excluded, viduals (97 female; mean age = 18.6 years) participated. r(108) = −.37, p < .001 for color; r(108) = −.33, p < .001 The data from 14 participants were removed on the basis for shape. Psychological Science 33(5) 827 Color Task Shape Task 10 10 0 0 −10 −10 0306090 120 0306090 120 HC Memory Precision (κ ) HC Memory Precision (κ ) M M Fig. 7. Results of the color task and shape task in Experiment 3. Each scatterplot depicts the relationship between bias for high-confidence (HC) reports and HC memory precision. Circles indicate individual participant data. Crosses indicate participant data excluded from the correlation as outliers. The solid lines indicate the best-fitting regressions. in different experimental blocks. We found that identi- General Discussion cal intervening inputs induced a larger memory bias in The present study revealed that the use of WM in per- compare blocks than in ignore blocks, and more impor- ceptual comparisons results in a systematic memory bias, tantly, we found that WM precision for the baseline thus demonstrating that WM is susceptible to interfer- condition (i.e., no intervening input) in both block ence even without task-irrelevant disruptions of active types was statistically indistinguishable. These results maintenance. This memory bias was particularly large demonstrated that the variability in WM precision alone when a visual input was subjectively similar to the WM. could not explain the memory bias and, thus, impli- Model comparisons showed that this similarity-induced cated a causal role of similarity judgments in modulat- memory bias was better characterized as representational ing the similarity-induced memory bias. integration between the WM and a subjectively similar The temporal stability of the similarity-induced mem- input (i.e., joint-density model) than probabilistic confu- ory bias also merits further discussion. Wang and col- sion between the two (i.e., mixture-density model). We leagues (2018) demonstrated that a WM representation also provided empirical support of this representational- can be distorted by similar perceptual inputs only when integration account by demonstrating that individuals they appear before attention returns to the WM from a with lower WM precision exhibited a larger memory bias secondary task. Although we did not use any secondary than those with higher precision. task to distract attention away from the original WM One may argue that the similarity-induced memory representation prior to the onset of the similar probe, bias is solely determined by WM precision, thus ques- it is possible that the similarity-induced memory bias tioning the causal role of perceptual comparisons in happens only when a WM is compared with a similar modulating the bias. More precisely, low-precision WM input too soon after encoding. We tested this possibility may be susceptible to greater stimulus-driven interfer- in a follow-up study (Saito et  al., 2021) by directly ence by a perceptual input, which may also result in manipulating the interval between memory encoding perceived similarity between them. In other words, per- and similarity judgments. Here, we replicated the ceived similarity between WM and a perceptual input similarity-induced memory bias even when participants may be a mere by-product of low WM precision. The performed a similarity judgment between a perceptual present study alone does not provide a strong test for input and a memory encoded 24 hr before by retrieving this possibility because we did not measure WM preci- it into WM (Atkinson & Shiffrin, 1968; Cowan, 1999; sion prior to the perceptual comparison. However, we Fukuda & Woodman, 2017). Furthermore, when par- tested this possibility in a follow-up study (Saito et al., ticipants retrieved a memory again 24 hr after a similar- 2020) by having participants either ignore or compare ity judgment, the memory reports remained biased. This intervening perceptual input with a WM representation temporal stability of the similarity-induced memory bias HC Bias (°) HC Bias (°) 828 Fukuda et al. This article has received the badge for Open Data. More is in stark contrast to the memory bias caused by a information about the Open Practices badges can be found temporary disruption of active WM maintenance, thus at http://www.psychologicalscience.org/publications/ suggesting that it is the direct usage of WM in percep- badges. tual comparisons that “locks in” the bias. Future studies will be necessary to determine the neural mechanisms that underlie these two distinct forms of memory biases. Our representational-integration account of WM bias Acknowledgments might offer a unifying explanation for other WM biases reported in the literature. For instance, past visual expe- We thank Steven J. Luck for the constructive comments on riences can bias subsequently encoded WM—a phe- the manuscript for this article. nomenon known as serial dependence in visual perception (Bae & Luck, 2019, 2020; Cicchini et  al., References 2018; Czoschke et al., 2019; C. Fischer et al., 2020; J. Atkinson, R. C., & Shiffrin, R. M. (1968). Human memory: Fischer & Whitney, 2014; Kiyonaga et  al., 2017). A proposed system and its control processes. In K. W. Although our finding that WM is biased by subsequently Spence & J. T. Spence (Eds.), The psychology of learning perceived inputs is different from serial dependence in and motivation: Advances in research and theory (Vol. 2, the direction of causality, serial dependence might also pp. 89–195). Academic Press. be explained by the representational integration trig- Bae, G.-Y. (2021). Neural evidence for categorical biases in location and orientation representations in a working gered by the subjective similarity between lingering memory task. NeuroImage, 240, Article 118366. https:// WM representations of past visual experience and cur- doi.org/10.1016/j.neuroimage.2021.118366 rent WM representations. Bae, G.-Y., & Luck, S. J. (2018). Dissociable decoding of Lastly, future studies should examine at what stage spatial attention and working memory from EEG the similarity-induced memory bias occurs. One pos- oscillations and sustained potentials. The Journal of sibility is that the memory bias might be introduced to Neuroscience, 38(2), 409–422. https://doi.org/10.1523/ WM as soon as its similarity to a novel input is per- JNEUROSCI.2860-17.2017 ceived. Alternatively, the WM might stay intact at the Bae, G.-Y., & Luck, S. J. (2019). Reactivation of previous time of the perceptual comparison, and the bias might experiences in a working memory task. Psycho logical be introduced at the time of memory report. To tease Science, 30(4), 587–595. https://doi.org/10.1177/095679 apart these two hypotheses, future studies should incor- 7619830398 Bae, G.-Y., & Luck, S. J. (2020). Serial dependence in vision: porate neural-decoding approaches (Bae, 2021; Bae & Merely encoding the previous-trial target is not enough. Luck, 2018; Rademaker et al., 2019) to track the content Psychonomic Bulletin & Review, 27(2), 293–300. https:// of WM as it is maintained, compared with a perceptual doi.org/10.3758/s13423-019-01678-7 input, and eventually reported. Bae, G.-Y., Olkkonen, M., Allred, S. R., & Flombaum, J. I. (2015). Why some colors appear more memorable than Transparency others: A model combining categories and particulars Action Editor: Barbara Knowlton in color working memory. Journal of Experimental Editor: Patricia J. Bauer Psychology: General, 144(4), 744–763. https://doi.org/10 Author Contributions .1037/xge0000076 K. Fukuda, A. E. Pereira, and H. Tsubomi designed the Bays, P. M., Catalao, R. F., & Husain, M. (2009). The precision study. A. E. Pereira and J. M. Saito collected the data. K. of visual working memory is set by allocation of a shared Fukuda analyzed the data. G.-Y. Bae conducted the model- resource. Journal of Vision, 9(10), Article 7. https://doi.org/ ing analysis. All the authors wrote the manuscript and 10.1167/9.10.7 approved the final manuscript for submission. Bennett, P. J., & Cortese, F. (1996). Masking of spatial fre- Declaration of Conflicting Interests quency in visual memory depends on distal, not retinal, The author(s) declared that there were no conflicts of frequency. Vision Research, 36(2), 233–238. https://doi interest with respect to the authorship or the publication .org/10.1016/0042-6989(95)00085-e of this article. Brainard, D. H. (1997). The Psychophysics Toolbox. Spatial Funding Vision, 10(4), 433–436. This research was supported by the Natural Sciences and Cicchini, G. M., Mikellidou, K., & Burr, D. C. (2018). The Engineering Research Council and the Connaught New functional role of serial dependence. Proceedings of the Researcher Award. Royal Society B: Biological Sciences, 285(1890), Article Open Practices 20181722. https://doi.org/10.1098/rspb.2018.1722 All data have been made publicly available via OSF and Cowan, N. (1999). An embedded-processes model of working can be accessed at https://osf.io/5v7sg/. The design and memory. In A. Miyake & P. Shah (Eds.), Models of working analysis plans for the experiments were not preregistered. memory: Mechanisms of active maintenance and executive Psychological Science 33(5) 829 control (pp. 62–101). Cambridge University Press. https:// Makovski, T., Sussman, R., & Jiang, Y. V. (2008). Orienting doi.org/10.1017/CBO9781139174909.006 attention in visual working memory reduces interference Czoschke, S., Fischer, C., Beitner, J., Kaiser, J., & Bledowski, from memory probes. Journal of Experimental Psychology: C. (2019). Two types of serial dependence in visual work- Learning, Memory, and Cognition, 34(2), 369–380. https:// ing memory. British Journal of Psychology, 110(2), 256– doi.org/10.1037/0278-7393.34.2.369 267. https://doi.org/10.1111/bjop.12349 Nemes, V. A., Parry, N. R., Whitaker, D., & McKeefry, D. J. Dubé, C., Zhou, F., Kahana, M. J., & Sekuler, R. (2014). (2012). The retention and disruption of color information Similarity-based distortion of visual short-term memory in human short-term visual memory. Journal of Vision, is due to perceptual averaging. Vision Research, 96, 8–16. 12(1), Article 26. https://doi.org/10.1167/12.1.26 https://doi.org/10.1016/j.visres.2013.12.016 Rademaker, R. L., Bloem, I. M., De Weerd, P., & Sack, A. T. Faul, F., Erdfelder, E., Lang, A.-G., & Buchner, A. (2007). (2015). The impact of interference on short-term memory G*Power 3: A flexible statistical power analysis program for visual orientation. Journal of Experimental Psychology: for the social, behavioral, and biomedical sciences. Human Perception and Performance, 41(6), 1650–1665. Behavior Research Methods, 39(2), 175–191. https://doi.org/10.1037/xhp0000110 Fischer, C., Czoschke, S., Peters, B., Rahm, B., Kaiser, J., & Rademaker, R. L., Chunharas, C., & Serences, J. T. (2019). Bledowski, C. (2020). Context information supports serial Coexisting representations of sensory and mnemonic infor- dependence of multiple visual objects across memory mation in human visual cortex. Nature Neuroscience, 22(8), episodes. Nature Communications, 11(1), Article 1932. 1336–1344. https://doi.org/10.1038/s41593-019-0428-x https://doi.org/10.1038/s41467-020-15874-w Rademaker, R. L., Park, Y. E., Sack, A. T., & Tong, F. (2018). Evidence Fischer, J., & Whitney, D. (2014). Serial dependence in visual of gradual loss of precision for simple features and complex perception. Nature Neuroscience, 17(5), 738–743. https:// objects in visual working memory. Journal of Experimental doi.org/10.1038/nn.3689 Psychology: Human Perception and Performance, 44(6), Fukuda, K., & Woodman, G. F. (2017). Visual working mem- 925–940. https://doi.org/10.1037/xhp0000491 ory buffers information retrieved from visual long-term Saito, J. M., Duncan, K., & Fukuda, K. (2021). Comparing memory. Proceedings of the National Academy of Sciences, visual memories to novel visual input risks lasting memory USA, 114(20), 5306–5311. https://doi.org/10.1073/pnas distortion. PsyArXiv. https://doi.org/10.31234/osf.io/xr4su .1617874114 Saito, J. M., Kolisnyk, M., & Fukuda, K. (2020). Perceptual Kiyonaga, A., Scimeca, J. M., Bliss, D. P., & Whitney, D. comparisons modulate memory biases induced by over- (2017). Serial dependence across perception, attention, lapping visual input. PsyArXiv. https://doi.org/10.31234/ and memory. Trends in Cognitive Sciences, 21(7), 493– osf.io/dqng3 497. https://doi.org/10.1016/j.tics.2017.04.011 Sun, S. Z., Fidalgo, C., Barense, M. D., Lee, A. C. H., Cant, J. S., Li, A. Y., Liang, J. C., Lee, A. C. H., & Barense, M. D. (2020). The & Ferber, S. (2017). Erasing and blurring memories: The validated circular shape space: Quantifying the visual simi- differential impact of interference on separate aspects of larity of shape. Journal of Experimental Psychology: General, forgetting. Journal of Experimental Psychology: General, 149(5), 949–966. https://doi.org/10.1037/xge0000693 146(11), 1606–1630. https://doi.org/10.1037/xge0000359 Luck, S. J., & Vogel, E. K. (2013). Visual working memory Teng, C., & Kravitz, D. J. (2019). Visual working memory capacity: From psychophysics and neurobiology to indi- directly alters perception. Nature Human Behaviour, 3(8), vidual differences. Trends in Cognitive Sciences, 17(8), 827–836. https://doi.org/10.1038/s41562-019-0640-4 391–400. https://doi.org/10.1016/j.tics.2013.06.006 Wang, B., Theeuwes, J., & Olivers, C. N. L. (2018). When Ma, W. J., Husain, M., & Bays, P. M. (2014). Changing con- shorter delays lead to worse memories: Task disruption cepts of working memory. Nature Neuroscience, 17(3), makes visual working memory temporarily vulnerable 347–356. https://doi.org/10.1038/nn.3655 to test interference. Journal of Experimental Psychology: Magnussen, S., & Greenlee, M. W. (1992). Retention and Learning, Memory, and Cognition, 44(5), 722–733. https:// disruption of motion information in visual short-term doi.org/10.1037/xlm0000468 memory. Journal of Experimental Psychology: Learning, Zhang, W., & Luck, S. J. (2008). Discrete fixed-resolution Memory, and Cognition, 18(1), 151–156. https://doi.org/ representations in visual working memory. Nature, 10.1037//0278-7393.18.1.151 453(7192), 233–235. https://doi.org/10.1038/nature06860

Journal

Psychological ScienceSAGE

Published: May 1, 2022

Keywords: working memory; memory distortion; individual differences; open data

There are no references for this article.