Color category and inter-item interaction influence color working memory codependently

Abstract

Our brains do not always encode visual information in a veridical way. Visual working memory (WM) for features such as color can be biased. WM bias comes from several sources. Category priors can lead to WM bias. For example, color WM is biased toward or away from category prototypes. In addition to category knowledge, contextual factors can induce and modulate WM bias; however, these biases of different sources have usually been investigated independently with different tasks. The present study sought to explore how color WM is influenced by both color category and concurrent distractor. Specifically, we asked participants to retain two color items in WM to investigate how the WM representation of the target color is biased by learned category knowledge and contextual inter-item interactions. Our study found that the WM representation of the target color is biased toward or away from the category prototypes and away from the distractor color that is simultaneously held in WM, indicating that both color category and concurrent distractor bias color WM. More importantly, the weight of these two biases depends on the specific color category, suggesting that category priors and inter-item interaction biases are not simply additive but flexible. Furthermore, we revealed that both types of biases arise from perceptual processes.

Introduction

Our brain does not always encode visual information veridically. For example, visual working memory (WM) for features such as color and orientation can be biased. WM bias derives from different sources. Category priors can induce WM bias. Color WM is reported biased toward or away from the category prototype (Bae, Olkkonen, Allred, & Flombaum, 2015; Spencer & Hund, 2002), and orientation WM is repelled from the cardinal axis (i.e., the cardinal bias) (Bae, 2022; Bae & Luck, 2019; Pratte, Park, Rademaker, & Tong, 2017; Wei & Stocker, 2015). In addition to innate or learned priors, contextual factors can also induce and modulate WM bias (Wedell, Hayes, & Kim, 2020). When multiple visual features are presented simultaneously or sequentially, they may interact with each other, giving rise to attraction or repulsion bias (Bae, 2022; Bae & Luck, 2017; Brady & Alvarez, 2011; Golomb, 2015; Huttenlocher, Hedges, & Duncan, 1991; Park & Zhang, 2024; Rademaker, Bloem, De Weerd, & Sack, 2015; Rauber & Treue, 1998; Rideaux & Edwards, 2016). In other words, the target WM is distorted by another present non-target feature. The WM bias caused by inter-item interaction is sensitive to other contextual variables, including target–distractor similarity (Foerster & Schneider, 2020), stimulus distribution (Foerster & Schneider, 2020; Hu, Wang, Talhelm, & Zhang, 2021; McKinley, Peterson, & Hout, 2023), and attentional allocation (Lee & Cho, 2019). 

The current study was interested in how color WM would be influenced by color category and concurrent distractor simultaneously. Specifically, we asked participants to maintain two color items in WM and explored how the WM representation of the target color was biased by category knowledge and contextual inter-item interaction. The impact of innate or learned priors and context in WM biases have been mostly investigated independently within different task settings (Esposito, Chiarella, Raffone, Nikolaev, & Van Leeuwen, 2023; Jabar & Fougnie, 2022; Jia, Cheng, Lu, Wu, & Wang, 2023; Langlois, Jacoby, Suchow, & Griffiths, 2021; Sampaio, Jones, Engelbertson, & Williams, 2020; Sreenivasan & Jha, 2007). On the one hand, researchers focusing on the effect of priors have often probed the WM bias of one single feature (Jabar & Fougnie, 2022; Langlois et al., 2021; Sampaio et al., 2020). On the other hand, most studies investigating contextual effects in WM bias have often ignored potential effects of priors (Esposito et al., 2023; Jia et al., 2023; Sreenivasan & Jha, 2007). 

Recently, a few attempts to explore how these two mechanisms operate simultaneously in WM bias (Bae, 2022; Taylor & Bays, 2018). Taylor and Bays (2018) found that the cardinal bias in orientation systematically decreases after exposure to a stimulus distribution that is incongruent to the orientation distribution in natural settings. This study suggested that the effect of priors in WM bias can be eliminated by contextual effects. Bae (2022) revealed similar results, that the cardinal bias in orientation was canceled out by inter-item interaction between sequentially presented orientations. However, the context effect caused by inter-item interaction in WM bias did not always win over the cardinal bias. Their relative weights in WM bias were modulated by attentional priority. In addition to those direct findings, the categorical bias in WM is in accordance with Bayesian theories of perception that the perception or recall of colors incorporates priors and generates biases. Numerous studies have shown that Bayesian priors can be adapted to changed stimulus environments occurring over short time scales (Adams, Rohlf, & Slice, 2004; Berniker, Voss, & Kording, 2010; Chalk, Seitz, & Series, 2010; Körding & Wolpert, 2004). Hence, we hypothesized that categorical bias might be outweighed by inter-item interaction in our study. 

Experiment 1

Methods

Participants

Twenty participants (eight males; mean age, 23.95 years old) participated in Experiment 1a, and 21 participants (nine males; mean age, 22.95 years old) participated in Experiment 1b. All participants reported having normal or corrected-to-normal vision. All of them had normal color vision. The participants provided consent before the experiment. The experiment was approved by the Ethics Committee at Beihang University, Beijing, China. 

Stimuli and procedure

All stimuli were presented on a 14-inch calibrated display (refresh rate, 60 Hz). The xyY values of the white point and the monitor primaries were measured. The input–output value of each channel was also measured to define the gamma curve. This information was used to determine the appropriate RGB values for each color stimulus as suggested by Brainard, Pelli, and Robson (2002). The experiment was run with MATLAB R2016a (MathWorks, Natick, MA) via Psychtoolbox (Brainard, 1997; Pelli, 1997). Participants sat at a viewing distance of 50 cm, such that the display subtended approximately 39.15° by 28.02° of visual angle. Each participant was required to complete two tasks in the following order: WM task and categorization task. 

WM task in Experiment 1a

The stimulus for the WM task was one colored disk. The surface color C1 was randomly selected from 23 colors (130°–240°; step size, 5°) from the green–blue range in the CIELAB space (L = 54; radius = 29) (Figure 1A), which only varied in hue. All of the color stimuli from the WM task were then used in the categorization task. In the WM task, each trial began with a fixation at the center of the screen for 1000 ms inside a rectangular area (Figure 1B). After that, one colored disk was presented in the center for 150 ms (presentation stage), followed by a 50-ms color mask (masking stage). The mask was a square with the same length as the diameter of the color stimulus (Figure 1C). The RGB values of each pixel of the mask were randomly and independently generated from a normal distribution with a mean of 128 and a standard deviation of 50. Then, a blank screen was presented for 750 ms (delay stage). After the delay stage, a color wheel was presented on the screen (response stage), and the participants clicked on the color wheel to report the hue that most closely resembled the remembered color. The participants could change their responses by clicking on different locations on the color wheel, and the hue clicked was displayed at the center. The color wheel was rotated randomly on each trial to prevent position–color association. There were 368 trials in total. 

Figure 1.

 

Stimuli and tasks in Experiment 1. (A) Hue ranges of two chromatic colors used for the WM task. (B) Procedure for the WM task. (C) Illustration of the color mask. (D) Procedure for the categorization task.

WM task in Experiment 1b

The stimuli for the WM task were two disks whose surface colors consisted of two chromatic colors (C1, C2). The color C1 was randomly selected from 23 colors (130°–240°; step size, 5°) from the green–blue range in the CIELAB space (L = 54; radius = 29) (Figure 1A), which only varied in hue. The color C2 was 15° (small distance) or 30° (large distance) larger than C1, thus ranging from 145° to 270°. The colors C1 and C2 had equal probabilities of being the target or the distractor color. In other words, the hue degree of the target color was larger than the distractor color on half the trials and smaller on the other half. All of the color stimuli from the WM task were then used in the categorization task. 

In the WM task, each trial began with a fixation at the center of the screen for 1000 ms inside a rectangular area (Figure 1B). After that, two disks with two surface colors C1 and C2 were presented at two locations for 150 ms (presentation stage), followed by a 50-ms mask (masking stage). Then, a blank screen was presented for 750 ms (delay stage). After the delay stage, one of the locations was given a cue for 50 ms instructing which of the two stimuli was the target (cueing stage). A color wheel was presented on the screen (response stage), and the participants clicked on the color wheel to report the hue that most closely resembled the remembered color. The participants could change their responses by clicking on different locations on the color wheel, and the hue clicked was displayed at the center. The color wheel was rotated randomly on each trial to prevent position–color association. There were 368 trials in total. Trials from small distance and large distance conditions were mixed and presented in a random order for each participant. 

Categorization task

In the categorization task, each trial began with a fixation at the center of the screen for 1000 ms. Then one colored disk was presented in the screen center. The participants were asked to make two judgments. First, they had to decide whether the color was “green” or “blue” by pressing “J” or “M” on the keyboard. After that, they had to rate how typical this color was as a category example by key presses on a five-point scale (1, not typical at all; 5, very typical). There were 174 trials in total (six trials for each color). 

Results

Category prototypes and boundary

Figures 2A and 2C plot the results of the categorization task in Experiments 1a and 1b. For each experiment, we first calculated the category boundary and prototypes at the group level. The average proportion of “blue” responses was fitted with a psychometric function defined as 1/(1 + exp(−(x − α)/β)), where α is the threshold (estimated value at which “blue” would be reported half of the time)—that is, the green–blue boundary (Figure 2A). The group boundary was 191.7° (95% confidence interval [CI], 190.7°–192.7°) in Experiment 1a, according to which we divided the color stimuli into two categories: green (145°–190°) and blue (195°–270°). The group boundary was 195.1° (95% CI, 194.4°–195.7°) in Experiment 1b (Figure 2C), according to which we divided the color stimuli into two categories: green (145°–195°) and blue (200°–270°). 

Figure 2.

 

Results of the categorization task in Experiments 1. (A) The green–blue boundary in Experiment 1a. (B) The location of the peak of the curve indicates the value of the color stimulus with the greatest typicality (i.e., the category prototype). The estimated prototypes were 144.0° (green) and 235.9° (blue) in Experiment 1a. (C) The green–blue boundary in Experiment 1b. (D) The estimated prototypes were 147.9° (green) and 234.0° (blue) in Experiment 1b.

To compute category prototypes, we fitted the average typicality ratings with von Mises functions: \(f( x ) = \frac{{\exp ( {k \times [ {\cos ( {x - \mu } )} ]} )}}{{[ {2\ \times \pi \times {{I}_0}( k ) \times m} ]}} + b\). The von Mises distribution is considered to be the circular analog of the normal distribution and thus was used because the color space is circular. In this model, the value of x corresponds to the location of the color stimuli in color space (i.e., hue degree). This function included four free parameters: µ (mean parameter), k (standard deviation parameter), m (modulation depth), and b (baseline). The mean parameter µ determines the location of the peak of the curve, and thus it indicates the value of the color stimulus with the greatest typicality (i.e., the category prototype). The model fitting was applied separately for each color category (green, blue) and each experiment. The estimated prototypes were 144.0° (95% CI, 141.9°–146.1°) for green and 235.9° for blue (95% CI, 234.4°–237.4°) in Experiment 1a (Figure 2B), and they were 147.9° (95% CI, 145.0°–150.9°) and 234.0° (95% CI, 233.1°–234.9°) in Experiment 1b (Figure 2D). Figures 2B and 2D plot the group-level boundary and typicality rating. Because there were individual differences in color categorization, there were infrequent cases where subjects classified a stimulus within the group-level boundaries of one color category as the other category (e.g., a “blue” classified as “green”). To give a clear picture of the individual differences, we additionally plot individual typicality ratings in the Supplementary Materials (Supplementary Figure S1). 

Category effects in WM bias in Experiment 1a

Figure 3A plots the distribution of mean reported errors (reported hue value minus target hue value) as a function of hue in Experiment 1a. We hypothesized that the color positions relative to the category prototypes (counterclockwise or clockwise) would influence the WM bias if the color WM is biased toward or away from the category prototypes. To test this category effect, we divided the color stimuli into four groups according to their category membership and positions relative to the category prototypes (Figure 3B): green colors counterclockwise to the prototype, green colors clockwise to the prototype, blue colors counterclockwise to the prototype, and blue colors clockwise to the prototype. As shown in Figure 3A, the colors located counterclockwise to the category prototypes were reported with greater positive errors than colors clockwise to the prototypes, suggesting that color WM is attracted toward the category prototypes. A 2 (color category: green or blue) × 2 (category position: counterclockwise or clockwise) repeated-measures analysis of variance (ANOVA) was conducted on the reported errors (Figure 3C). We found a significant interaction effect, F(1, 19) = 22.44, p < 0.001, partial η² = 0.542, and significant main effects of color category, F(1, 19) = 10.43, p = 0.004, partial η² = 0.354, and category position, F(1, 19) = 163.16, p < 0.001, partial η² = 0.896. Further simple-effect analyses revealed that the category position significantly influenced the WM bias for both green and blue colors: green, t(19) = 5.14, p < 0.001, Cohen's d = 1.150; blue, t(19) = 12.24, p < 0.001, Cohen's d = 2.737. The effect of category position was more pronounced for blue colors than green colors, which is probably associated with the uneven distribution of color ranges across color categories. The green category had a wider color range, and the blue category had a narrower color range. Consequently, the WM bias for green colors changed less drastically as a function of hue than for blue colors. 

Figure 3.

 

Results for Experiment 1. (A) The distribution of mean reported errors (reported hue minus target) as a function of hue in Experiment 1a. (B) Condition assignment, where the color stimuli were divided into four groups based on their category membership and position relative to the category prototypes. (C) Category effect in Experiment 1a. (D) The distribution of mean reported errors (reported hue value minus target value) as a function of hue value in Experiment 1b. (E) Category effect in Experiment 1b. (F) Inter-item effect in Experiment 1b. *p < 0.05, **p < 0.01, ***p < 0.001.

Category and inter-item effects in WM bias in Experiment 1b

We first examined whether and how color category and inter-item interaction influence WM bias independently in Experiment 1b. Figure 3D plots the distribution of mean reported errors (reported hue value minus target hue value) as a function of hue. As in Experiment 1a, we divided the color stimuli into four groups according to their category membership and positions relative to the category prototypes (Figure 3B). Consistent with the findings in Experiment 1a, the colors located counterclockwise to the category prototypes were reported with greater positive errors than colors located clockwise to the prototypes, suggesting that color WM is attracted toward the category prototypes. However, the category effects in Experiment 1b were less obvious compared to Experiment 1a, because the mean reported errors were more uniformly distributed as a function of hue value. A 2 (color category: green or blue) × 2 (category position: counterclockwise or clockwise) repeated-measures ANOVA was conducted on the reported errors (Figure 3E). We found a significant interaction effect, F(1, 20) = 11.713, p = 0.003, partial η² = 0.369, and main effect of position, F(1, 20) = 24.44, p < 0.001, partial η² = 0.550. The main effect of color category was not significant, F(1, 20) = 0.81, p = 0.38, partial η² = 0.039. Further simple-effect analyses revealed that the category position significantly influenced WM bias for blue colors, t(20) = 5.98, p < 0.001, Cohen's d = 1.304, but not green colors, t(20) = 1.82, p = 0.083, Cohen's d = 0.398. 

Figure 3F plots the mean reported errors when the relative position of the target color to the distractor color on the green–blue scale was varied (counterclockwise to distractor vs. clockwise to distractor). Note that leftward positions mean smaller hue values and rightward positions mean larger hue values. The target color was reported with greater positive errors when the target was clockwise to the distractor compared to when the target was counterclockwise to the distractor. Paired t-tests revealed significant differences between these two conditions, t(20) = 3.05, p = 0.006, Cohen's d = 0.666. This finding suggested that these two concurrent items repel each other. 

Co-effects of color category and inter-item interaction in Experiment 1b

We then examined the co-effects of color category and inter-item interaction in WM bias (Figures 4A and 4B). A 2 (color category: green or blue) × 2 (category position: counterclockwise to prototype or clockwise to prototype) × 2 (target–distractor position: target counterclockwise to distractor or target clockwise to distractor) repeated-measures ANOVA was conducted on the reported errors, which revealed a significant three-way interaction, F(1, 20) = 13.36, p = 0.002, partial η² = 0.401. Hence, we tested the effects of category position and target–distractor position for green and blue colors separately. For green colors, the repeated-measures ANOVA (category position: counterclockwise to prototype or clockwise to prototype; target–distractor position: target counterclockwise to distractor or target clockwise to distractor) suggested a significant interaction, F(1, 20) = 9.71, p = 0.005, partial η² = 0.327. Further analyses revealed that category position significantly influenced color WM when the target color was counterclockwise to the distractor (p < 0.001), but not the other way around (p = 0.256). Target–distractor position significantly influenced color WM regardless of the category position of the target (p < 0.01). For blue colors, repeated-measures ANOVA (category position: counterclockwise to prototype or clockwise to prototype; target–distractor position: target counterclockwise to distractor or target clockwise to distractor) also suggested a significant interaction, F(1, 20) = 7.77, p = 0.011, partial η² = 0.280. Further analyses revealed that category position significantly influenced color WM regardless of the target–distractor position (p < 0.001). Target–distractor position significantly influenced color WM when the target was counterclockwise to the category prototype (p = 0.005) but not when it was clockwise to the prototype (p = 0.203). We believe that the differential effects of category position and target–distractor position can be attributed to the stimulus-specific variations in category effects (Bae et al., 2015). Inter-item interaction could be less pronounced for the color stimuli with great category bias and more obvious for those with small category bias. In summary, these results suggest that color category and inter-item interaction do not influence color WM in parallel but codependently. 

Figure 4.

 

The co-effects of color category and inter-item interaction in WM bias in Experiment 1b. (A) The distribution of mean reported errors (reported hue minus target) as a function of hue when the target–distractor position was varied. (B) Category and inter-item effects for the green and blue category. The black arrows indicate the direction of the category effect (leftward means clockwise) for the target color, and the orange arrows indicate the direction of the inter-item effect. *p < 0.05, **p < 0.01, ***p < 0.001.

Additional analyses: Central tendency bias and serial bias in Experiment 1

Our study used a limited sample space, which created a risk of biased responses, such as the central tendency bias (i.e., regression toward the mean). To test whether there was a central tendency bias and how it may interact with the bias of interest (i.e., category and inter-item bias), we analyzed the error data from Experiment 1b using multiple linear regression. We first built a model (termed Model A) with category position (counterclockwise to prototype or clockwise to prototype) and target–distractor position (target counterclockwise to distractor or target clockwise to distractor) as predictive variables. The data for all participants were pooled together. On the basis of Model A, we constructed a new model with an additional variable of actual color value (termed Model A.1). We assumed that a central tendency bias would be manifested by a negative correlation between the actual color value and the reported errors (Olkkonen, McCarthy, & Allred, 2014). In the presence of a central tendency, a significant main effect of the actual color value was anticipated, with Model A.1 outperforming Model A.1. The results of regression analysis (Supplementary Tables S1 and S2) showed that Model A was statistically significant, F(2, 7725) = 189.63, p < 0.001, and accounted for 4.7% of the variance in reported error (R² = 0.047). Both category position (p < 0.001) and target–distractor position (p < 0.001) were found to be significant predictors of reported error. Model A.1 was found to be statistically significant, F(3, 7724) = 169.43, p < 0.001, and accounted for 6.1% of the variance in reported error (R² = 0.061). All variables, including category position (p < 0.001), target–distractor position (p < 0.001), and actual color value (p < 0.001), were found to be significant predictors of reported error. The results from Experiment 1b suggest that the actual color value negatively predicts the reported error of the target, indicating the presence of central tendency bias. However, the presence of central tendency bias did not influence the main results of our study. After including the variable of actual color value in the model, we still observed significant category and inter-item effects. 

We additionally investigated serial dependence effects in our study. To probe the presence of serial dependence effects, we tested whether the target representation in the current trial was attracted to or repulsed by the target color in the previous trial. We compared the mean reported errors when the target color in the previous trial was clockwise versus counterclockwise to the target color in the current trial. As shown in Supplementary Figure S2, the target color was reported with greater positive errors when the target color in the current trial was clockwise (i.e., smaller hue value) to the target color in the previous trial compared with when the target color in the current trial was counterclockwise (i.e., larger hue value) to the target color in the previous trial, indicating attraction bias in both experiments. Repeated-measures ANOVA (current or previous position: clockwise or counterclockwise; experiment: Experiment 1a or Experiment 1b) revealed a significant effect of current–previous position, F(1, 39) = 14.89, p < 0.001, partial η² = 0.276. Neither the effect of experiment, F(1, 39) = 0.014, p = 0.908, partial η² = 0.000, nor interaction, F(1, 39) = 3.27, p = 0.078, partial η² = 0.077, was significant. These results suggest that the target representation in the current trial was attracted by the target color in the previous trial, no matter whether the target color was concurrent with another color (Experiment 1b) or not (Experiment 1a). 

Discussion

Experiment 1 suggested that the WM representation of the target color is biased toward category prototypes and away from the distractor color simultaneously held in WM, indicating the presence of categorical biases and inter-item biases in color WM. Furthermore, we found that the weights of these two biases depend on specific color categories. For green colors, the bias caused by inter-item interaction was persistent regardless of the category position of the target, whereas the category bias was only observed when the target was pushed counterclockwise by the distractor. For blue colors, the category bias was persistent regardless of the position of the target relative to the distractor, whereas the bias caused by inter-item interaction was only observed when the target was counterclockwise to the category prototype. The additional analysis showed the presence of a central tendency bias and a serial bias in Experiment 1. The target representation in the current trial was drawn to the target color in the previous trial and regressed toward the mean color value at the same time, but these biases have not influenced our main results. One may wonder whether categorical biases and inter-item biases would generalize to other color categories because we used only a limited number of color samples in Experiment 1. Hence, we conducted Experiment 2 where the purple–pink range was used for stimulus sampling. In addition, Experiment 2 was also intended to investigate whether these effects derive from a perceptual or mnemonic process by implementing both delayed and undelayed estimation procedures in the WM task. 

Experiment 2

The aim of Experiment 2 was twofold. First, we intended to examine whether the category and inter-item effects observed in Experiment 1 would generalize to other color categories. Second, Experiment 2 aimed to investigating whether these effects derive from a perceptual or mnemonic process. In this experiment, we used a purple–pink range for stimulus sampling and employed both delayed and undelayed estimation procedures (Bae et al., 2015). We hypothesized that the effect would be more exaggerated with a delayed estimation procedure than an undelayed estimation procedure if the effects had mnemonic origins. 

Methods

Participants

Twenty-one participants (13 males; mean age, 23.67 years old) participated in Experiment 2. All participants reported having normal or corrected-to-normal vision. All of them had normal color vision. The participants provided consent before the experiment. The experiment was approved by the Ethics Committee at Beihang University, Beijing, China. 

Stimuli and procedure

The stimuli and procedure were identical to Experiment 1b except for the following changes. The color C1 was randomly selected from among 26 colors (270°–345°; step size, 3°) from the purple–pink range in the CIELAB space (L = 54; radius = 29) (Figure 1A) and only varied in hue. The color C2 was 30° larger than C1, thus ranging from 300° to 375°. As in Bae et al. (2015), we implemented both delayed and undelayed estimation procedures in the WM task. For delayed estimation, there was a delay of 1000 ms after the mask presentation. For undelayed estimation, the color wheel for response was presented immediately after the mask. There were two blocks for each estimation procedure, and the order of the two estimation procedures was balanced across participants. Each participant completed four blocks of 156 trials, totaling 624 trials. The categorization procedure was identical to Experiment 1 except that the participants decided whether the color was “purple” or “pink” rather than “green” or “blue.” 

Results

Category prototypes and boundary

As in Experiment 1, we first calculated category boundary and prototypes at the group level in Experiment 2 (Figures 5A and 5B). The group boundary was 325.4° (95% CI, 290.9°–353.9°), according to which we divided the color stimuli into two categories: purple (270°–324°) and pink (327°–375°). The estimated purple prototype was 310.4° (95% CI, 309.2°–311.7°), and the estimated pink prototype was 351.6° (95% CI, 349.8°–353.3°). 

Figure 5.

 

Results in Experiment 2. (A) The purple–pink boundary. (B) The location of the peak of the curve indicates the value of the color stimulus with the greatest typicality (i.e., the category prototype). The estimated prototypes are 314.0° (purple) and 351.6° (pink). (C) The distribution of mean reported errors (reported hue minus target) as a function of hue. (D) Category effect. (E) Inter-item effect in Experiment 1b. (F) The co-effects of color category and inter-item interaction in WM bias for delayed and undelayed estimation. *p < 0.05, **p < 0.01, ***p < 0.001.

Category and inter-item effects for purple and pink

Figure 5C plots the distribution of mean reported errors (reported hue value minus target hue value) as a function of hue in Experiment 2. As in Experiment 1, we divided the color stimuli into two groups according to their category membership and positions relative to the category prototypes: purple colors clockwise to the prototype, purple colors counterclockwise to the prototype, pink colors clockwise to the prototype, pink colors counterclockwise to the prototype. Figure 5D plots the mean reported errors of the target color for two categories when category position was varied. The pattern of the category effect was distinct from that in Experiment 1. The colors located clockwise to the category prototypes were reported with greater positive errors than colors located counterclockwise to the prototypes, suggesting that color WM is pushed away from the category prototypes. We conducted a 2 (color category: purple or pink) × 2 (category position: clockwise to prototype or counterclockwise to prototype) repeated-measures ANOVA on the reported errors. We found a significant interaction effect, F(1, 20) = 14.59, p < 0.001, partial η² = 0.422, and significant main effects of color category, F(1, 20) = 14.00, p < 0.001, partial η² = 0.412, and category position, F(1, 20) = 35.80, p < 0.001, partial η² = 0.642. Further simple-effect analyses revealed that the category position significantly influenced WM bias for purple colors but not pink colors: purple, t(20) = 6.74, p < 0.001, Cohen's d = 1.471; pink, t(20) = 0.361, p = 0.722, Cohen's d = 0.079. 

The pattern of inter-item effect was consistent with that observed in Experiment 1 (Figure 5E). The target color was reported with greater positive errors when the target was clockwise to the distractor compared with when the target was counterclockwise to the distractor. Paired t-tests revealed significant differences between these two conditions, t(20) = 11.73, p < 0.001, Cohen's d = 2.561. It suggested that these two concurrent items repel each other. 

Co-effects of color category and inter-item interaction: Delayed versus undelayed estimation

Then we investigated whether categorical and inter-item effects would be influenced by delayed versus undelayed estimation (Figure 5F). After pooling two categories together, we conducted a 2 (delay: delayed or undelayed) × 2 (category position: clockwise to prototype or counterclockwise to prototype) × 2 (target–distractor position: target counterclockwise to distractor or target clockwise to distractor) repeated-measures ANOVA on the reported mean errors. We found a significant interaction effect between category position and target–distractor position, F(1, 20) = 21.41, p < 0.001, partial η² = 0.517, replicating the findings in Experiment 1b. Target–distractor position also significantly influenced the reported error of the target, F(1, 20) = 130.39, p < 0.001, partial η² = 0.867, but not category position, F(1, 20) = 0.03, p = 0.859, partial η² = 0.002. Further analyses revealed that category position significantly influenced color WM regardless of the target–distractor position (p < 0.01). Target–distractor position significantly influenced color WM regardless of the category position of the target (p < 0.001). The interaction between category position and target–distractor position was mainly caused by a greater inter-item effect for the target colors that were counterclockwise to the prototype than for those clockwise to the prototype. 

Importantly, we found no evidence of delay effects. The main effect of delay was not significant, and the variable of delay did not interact with other variables (p > 0.1). In other words, none of the category effects, inter-item effects, or their interaction was modulated by whether the recall was immediate or delayed, suggesting perceptual origins for these effects. The finding that category effects are not modulated by delay is surprising, as Bae et al. (2015) revealed greater category bias for delayed estimation than undelayed estimation. We suspected that the null effect of delay on category effect could be associated with the insignificant category biases in the pink category observed earlier. Hence, we assessed the effect of delay separately for the purple and pink categories. The category effect for each color category was quantified by the differences in reported errors between the two category position conditions (counterclockwise to prototype minus clockwise to prototype). We found that, for the purple category, delayed recall induced marginally significantly greater category biases than undelayed recall, t(20) = 2.04, p = 0.055, Cohen's d = 0.444. No difference was found for the pink category between the two delay conditions, t(20) = 0.69, p = 0.496, Cohen's d = 0.151. Even though there is some evidence that the category effect was modulated by immediate versus delayed recall, it is noteworthy that the category effect was present for undelayed estimation, in favor of a perceptual origin. 

Discussion

We found in Experiment 2 that inter-item biases were not limited to green and blue categories and can be generalized to other color categories. Interestingly, we revealed category bias for the purple category but not the pink category. The insignificant category effect for the pink category might be associated with the category typicality of the color samples used in this experiment. As is shown in Figure 5B, the ratings of category typicality of the pink colors are more homogeneously distributed, compared to colors of other categories (e.g., purple, green). Furthermore, Experiment 2 revealed that the category and inter-item biases may arise from perceptual processes instead of mnemonic processes. However, the category bias, present for undelayed estimation, is also influenced by whether or not the recall is delayed. For the purple category, delayed recall induced greater category biases than undelayed recall, consistent with Bae et al. (2015). 

General discussion

Results from our study suggested that the WM representation of the target color is biased toward category prototypes and away from the distractor color simultaneously held in WM, indicating both categorical biases and inter-item biases in color WM. More importantly, these two types of bias interact with each other, and the weights of these two biases depend on the specific color category of the target, as well as the position within its category. Furthermore, we found that the inter-item biases may arise from perceptual processes instead of mnemonic processes because the magnitudes of the inter-item biases were comparable for the delayed and undelayed estimations. In contrast, the category bias appeared with undelayed recall but increased when the recall was delayed. 

In contrast to previous studies that separately investigated categorical bias and inter-item interaction in color WM (Esposito et al., 2023; Jabar & Fougnie, 2022; Jia et al., 2023; Langlois et al., 2021; Sampaio et al., 2020; Sreenivasan & Jha, 2007), the current study is the first, to our knowledge, to explore their interactions. First, our results suggested the emergence of both biases when two items are held in WM, and these biases are not limited to specific color categories. The target color is attracted toward (e.g., green) or repelled from (e.g., purple) category prototypes. This finding was consistent with previous studies that have reported similar systematic biases for color and orientation WM due to learned priors (Bae et al., 2015; Bae & Luck, 2019; Pratte et al., 2017; Wei & Stocker, 2015). In addition to the category effect, we also found a significant inter-item interaction effect such that the two color items in WM repel each other. This finding is in accordance with some previous studies on the inter-item interaction effect (Bae & Luck, 2017; Golomb, 2015; Rademaker et al., 2015; Rauber & Treue, 1998; Rideaux & Edwards, 2016). However, previous research has found inter-item interactions in the form of attraction (Brady & Alvarez, 2011; Huttenlocher et al., 1991; Park & Zhang, 2024) instead of the repulsion present in our study. This discrepancy can be explained by the view of memory fidelity (Chunharas, Rademaker, Brady, & Serences, 2022; Lively, Robinson, & Benjamin, 2021). WM items with high memory fidelity tend to repel representations of similar items, and those with low memory fidelity tend to attract representations of similar items (Chunharas et al., 2022; Lively et al., 2021). As this study presented only two color items for memory, the participants were able to effectively process the color information. Consequently, the WM representation of the color items had relatively high memory fidelity, resulting in a repulsion effect during inter-item interaction. 

More importantly, our study suggested that these two effects are not simply additive but rather flexibly interact with one another. The interaction between category and inter-item effects depends on the specific color category of the target and the specific position within its color category. Bayesian theories of perception suggest that the perception or recall of colors incorporates priors and produces perceptual biases and that the weight of the prior is adjusted according to the contextual content. Numerous studies have shown that Bayesian priors can be adapted to changed stimulus environments occurring over short time scales (Adams et al., 2004; Berniker et al., 2010; Chalk et al., 2010; Körding & Wolpert, 2004). These theories can explain our finding that the weight of the category priors was not fixed but varied depending on the concurrent memoranda. Furthermore, the results of Experiment 2 revealed that both categorical and inter-item effects may arise from perceptual processes. Brouwer and Heeger (2013) proposed that the fine-grained neural color space in visual cortex can be warped into a categorical space during active color categorization. This categorical clustering is attributed to a color-specific gain change, and the gain of each neuron changes as a function of its selectivity relative to the category prototype. This categorical space may be induced not only by active color categorization but also by color estimation, delayed or not (Bae et al., 2015). This explains why we observe category bias for both immediate and delayed estimation. When there is a concurrent color representation, inter-item bias is superimposed on category bias. The addition of these two effects contains non-uniformities and is modulated by the specificity of the target color within the feature space (Schurgin, Wixted, & Brady, 2020). The inter-item bias can be explained by the underlying neural mechanisms of lateral inhibition (Blakemore, Carpenter, & Georgeson, 1970). The lateral inhibition explanations propose that visual features (like color) may be represented in a map-like manner, where neighboring neurons encode neighboring parts of the feature space. Lateral inhibitory connections between these neurons help to enhance the precision of feature representations. Consequently, if similar features are present, and their representations are coded by adjacent neurons, these neurons may inhibit each other. As a result, the neurons representing this similar region of feature space become relatively suppressed, leading to a bias that pushes the feature representations apart. 

There is some evidence that the category bias is modulated by delayed or undelayed estimation. Delayed recall leads to greater category bias than undelayed recall for the purple category but not the pink category. The finding that category bias increased under delayed recall is consistent with previous findings (Bae et al., 2015; Yan, Christophel, Allefeld, & Haynes, 2023). Even though the category bias emerges for immediate recall, an extended duration leads to a more categorical encoding of the stimuli (Yan et al., 2023). In addition to the biases of interest, the color WM was subject to other biases in our study. We revealed the emergence of a central tendency bias and a serial dependence bias. Specifically, the target representation in the current trial was attracted to the target color in the previous trial and biased toward the mean color value in the meantime. Also, we observed that the target representation in Experiment 1a and 1b were positively biased overall but not in Experiment 2 (Supplementary Figure S3). This difference may lie in the distinct color sampling procedures between these two experiments. In Experiment 1, hues were sampled from the blue–green part of the CIELAB space, whereas the purple–pink part was sampled in Experiment 2. The positive bias in hue estimation was also reported in a previous study (Experiment 2 in Olkkonen et al., 2014) in which the blue–green range was used. However, these biases were not effects of interest in our study and did not influence our main results. Future work may choose to quantify the biases of different sources, including central tendency bias and serial dependence. 

Acknowledgments

Supported by the Natural Science Foundation of Jiangsu Province (BK20230489) and Social Science Foundation of Jiangsu Province (23YYC005). 

Commercial relationships: none. 

Corresponding authors: Mengdan Sun; Chundi Wang. 

Emails: [e-mail]; [e-mail]. 

Address: Department of Psychology, Soochow University, Suzhou, China; Department of Psychology, School of Humanities and Social Sciences, Beihang University, Beijing, China.