Category relevance attenuates overshadowing in human predictive learning.

Experiment 1

Experiment 1 was run to assess how participants learn to predict an outcome using a color cue, a symbol cue, and a color–symbol compound in a within-subjects design (see Table 1 ). We anticipated that the cues trained alone will acquire higher predictive value than those trained as part of a compound (i.e., overshadowing effect). Stimuli belonging to both categories (i.e., compound) were presented simultaneously and formed a unique cue, with the symbol displayed over the color in the same physical space (see Figure 1A ). The first category, symbol, was made of black Chinese characters; the second one was made of different colors. We anticipated that colors would be easier to process than the complex symbols (participants with no prior knowledge of Chinese were recruited). Colors are used in all the aspects of our daily life and they are processed automatically, requiring little effort to process different colors and discriminate between them. The colors used in our task were highly discriminable (without subtle differences along a specific category of the color like hue or luminance). Moreover, there is evidence that color is the most important category controlling behavior of nonhuman animals and humans in biologically relevant tasks ( Kazemi et al., 2014 ; Sherratt et al., 2015 ). In terms of learning, for example, pigeons trained to predict outcomes based on colors and orientation lines showed an advantage to learn cues in the color category ( Mackintosh & Little, 1969 ). On the contrary, complex symbols are more demanding, forcing participants to focus on specific details to discriminate between them. Such differences presumably result in two separable categories independent from each other. Control cues for each category (color and symbol) were also trained, enabling us to simultaneously assess overshadowing in both categories (see Table 1 ).

In addition to using two categories, color and symbol, with different salience or effectiveness (as argued above), we also explored whether the length of exposure to the cues affects the magnitude of the overshadowing effect. Three groups of participants were given short (1s), medium (3s), or long (9s) exposure to the relevant cues (colors, symbols, and color–symbol compound). We conjectured that the amount of time exposure to the compound would determine the way in which participants integrate the information relative to each element of the compound. With a short exposure, participants may just focus on determining the relationship of the most discriminative and salient category, the color, neglecting the cue belonging to the alternative category, the symbol. However, longer exposure may allow adequate processing of both categories, allowing the integration of the symbol cues. Previous work suggests that long stimuli attenuate overshadowing in rats ( Sissons et al., 2009 ) and reduce overselectivity in humans ( Reynolds & Reed, 2018 ). However, we are not aware of studies exploring whether the signal length is a factor determining overshadowing in humans. In a recent experimental series from our laboratory ( Herrera et al., 2022 ), we did not observe reduced overshadowing with longer CSs; however, this was based on a comparison across experiments, so it should not be considered as conclusive evidence for the absence of an effect of stimulus length in overshadowing.

Assuming that competition is the default outcome when facing a compound of two stimuli, an overshadowing effect was expected. However, with short exposure, given the anticipated salience imbalance, the color was expected to be more effective in overshadowing the symbol than the other way around. On the other hand, long exposure to the compound might allow a more efficient processing of the symbol category, contributing to equate the salience of the color and symbol categories. With similarly high salient competing stimuli, an attenuation of overshadowing could be expected ( Mackintosh, 1976 , Experiment 2). As an alternative, long exposure might allow for the processing of the compound as a configuration. A bias to encode the compound configurally in the long-exposure group (9s) might therefore attenuate cue competition ( Urcelay & Miller, 2009 ; Williams et al., 1994 ).

In line with what Lovibond et al. (2020) have done in generalization experiments, we were also interested in the spontaneous strategies declared by participants after training with the compound color–symbol cues. In an overshadowing paradigm, the two elements of the compound CS have in principle equal opportunities to become associated with the outcome, although as mentioned above, this is influenced by the salience of the stimuli ( Mackintosh, 1976 ). However, participants might develop a spontaneous preference for one of the elements of the compound and attend exclusively to that category—increasing its salience. Alternatively, they could try to process both elements of the compound, which would lead to the development of predictive value by both elements and, therefore, an increase in their salience. To the best of our knowledge, the spontaneous strategy declared by participants following the completion of the task has not received much attention in compound training phenomena such as the overshadowing effect. To evaluate the strategy declared by participants, we included a forced-choice question at the end of the experiment (see Lovibond et al., 2020 ). If participants declared only processing one category (either the color or the symbol), we may expect an asymmetrical overshadowing, since the spontaneous preference for the color might prevent the processing of the alternative category (symbol). In that case, given that the symbol is ignored, the color would enter into association without competition and no difference should be expected with the color trained by itself (the control condition): the preferred category (the color) overshadows the alternative category, but not the other way around. The same might

Design

A 3 (group: 1s, 3s, 9s) × 2 (category: color vs. symbol) × 2 (cue: control vs. target) mixed-design was used. The group factor was manipulated between-subjects, and each group experienced the stimuli for a fixed amount of time, either 1s, 3s, or 9s. The other factors were manipulated within-subjects. For the factor category, there were two categories of stimuli: color and symbol. Color category was represented by a square of several colors and the symbol category by using different Chinese characters. The factor cue had two levels: control refers to the cue trained alone (in both categories) and target to the cue trained in compound with a cue of the alternative category. For example, as the upper part of Table 1 shows, for the category color, the control cue A was trained alone and the critical comparison is with the target cue C (trained in compound with cue X—a symbol). Note that this design allows a control cue in each category (A and V) to be compared with the corresponding target cue in each category trained in compound (CX): the color A is compared to the color C whereas the symbol V is compared to the symbol X. The same applied to the nonreinforced cues, trained alone (B, W) or as part of a compound (DY).

Materials and Apparatus

The task was inspired by Lovibond et al.’s (2020) task. The experiment was programmed and hosted online using the Gorilla Experiment Builder ( Anwyl-Irvine et al., 2020 ). Two set of stimuli were used, colors and symbols. A colored square ( Figure 1A ) of approximately 300 × 300 pixels and a black Chinese character (see Figure 1B ) of approximately 200 × 200 pixels over a white background were used as stimuli. When stimuli appeared as a compound, the symbol was presented over the colored square (see Figure 1A ). As colors we used: green (RGB: 90, 197, 58), blue (RGB: 56, 128, 247), orange (RGB: 245, 195, 66), pink (RGB: 234, 51, 157), and purple (RGB: 104, 52, 154). As symbols we used the characters for animals: bird (鸟), donkey (驴), squirrel (松), raccoon (狸), and kangaroo (袋). These stimuli were counterbalanced across subjects.

Procedure

After reading and signing the consent form, participants were presented with the following instructions:

[Screen 1] Please, read the instructions for the task carefully. After reading the instructions, you will be asked two questions to make sure you have understood them. If you fail to respond correctly to all questions, you will have the chance to read the instructions again. You will not be able to start the experiment until you respond correctly to all questions.

[Screen 2] We would like you to imagine that you have come across a strange machine. It appears to have a display on it, as well as poster that says “WARNING: this machine gives electric shocks!! When you see warning signs like this ______, do NOT touch!”

Unfortunately, the area of the label that shows the warning signals has been scratched off, so you do not know which signs predict danger. Your job is to work out what kinds of signs on the machine predict the shock.

[Screen 3] The experiment is composed of a number of “trials.” On each trial, you will be presented with a sign on the shock machine. You will then make a prediction about whether you think a shock will occur. In the FIRST phase of the experiment, you will receive feedback for your predictions about whether a shock occurred or not. In a final phase, you will need to rate on a scale to what extent you think each sign caused a shock, but you are not going to receive feedback. You will receive further instructions at the beginning of this phase.

[Screen 4] In the first phase, you will learn which signs lead to shock. We will present the same signs to you MULTIPLE times. On each trial, the sign will appear on the screen and next, the question “The sign above appears on the machine. What do you think will happen?” Press “z” if you think NO SHOCK will occur, or “m” if you think a SHOCK will occur. You need to wait for the question to appear to make your decision. From the moment the question appears you need to respond quickly. You will have 1.5 s to respond. You must respond within this time. If you do not respond during this time, you will receive feedback, however, your response will not be recorded. Look at the signs and the feedback carefully. Use this feedback to find out which signs lead to shock. Don´t worry, at first you will have to guess because you do not know much about these signs, but eventually you will learn which sign leads to shock and you will be able to make the correct predictions. If you are not able to respond, you still will receive feedback, however, your response will not be recorded. Please, try to avoid to the best of your abilities trials without a response.

After the instructions, a couple of forced-choice questions were introduced to check whether participants understood the instructions. In case participants failed any of the questions, they had to read the instructions again until they correctly responded to both questions. The questions and answers were: 1 What are you instructed to do during the task?

To evaluate whether or not different signs predict a shock.

To respond as fast as possible when you see a sign.

To evaluate how beautiful are the signs to you.

What is the key to select “no shock?”

Once participants had correctly responded to these questions, the training phase started. On each trial, the stimuli appeared for 1s/3s/9s depending on the experimental group (see Figure 1B ). After the respective time for each group had elapsed, the question “What do you think will happen?” NO SHOCK press “Z” or SHOCK press “M” appeared on the screen (the stimuli remained on the screen). The question and stimuli were presented either until the participant responded or for a maximum of 1,500 ms—the time window to register the response. Following the response or time limit the screen displayed corrective feedback (“correct”; “incorrect”; or “too slow” in those trials in which the participant failed to respond within the 1,500 ms time window) and information on whether the outcome was present (“the previous sign produces a shock”—with an image of a virtual shock) or absent (“the previous sign did not produce a shock”). Feedback was displayed for 3s in the absence of the stimuli, followed by a 1,000 ms intertrial interval (ITI) during which a white screen was displayed. The ITI was kept constant in all groups to avoid benefits in terms of spacing of trials.

The training phase consisted of eight presentations of each cue (see Table 1 ). Training was divided into four blocks of 16 trials with two presentations of each cue per block (64 trials in total). The order of trials within each block was randomly determined without any restrictions. For the color category, two cues were presented alone, one reinforced (A+, the control cue), and another not reinforced (B−). Similarly, for the symbol category, two cues were presented alone, one reinforced (V+, the control cue), and another not reinforced (W−). Two additional colors (C and D) were presented in a reinforced compound (CX+; C was the target cue for the color category and X for the symbol category) and as a nonreinforced compound (DY−). Reinforced and nonreinforced cues were always followed or not by the outcome according to their programmed contingency. The partial color (E±) and the partial symbol (Z±) cues received one reinforced trial and one nonreinforced trial per block. The partial cues were introduced to introduce some variability to avoid participants attention wandering during the task.

During the test phase, a single test question was presented with an image of each cue alone (each letter in Table 1 ). Participants read the following question: “What is the likelihood of the sign above leading to SHOCK?” There was a horizontal rating scale, ranging from 0 ( definitely no shock ) to 100 ( definitely shock ). The pointer was positioned at 50 on the scale, and participants could move the pointer in either direction with the mouse. There was no time limit to respond to each qu

Transparency and Openness

We report how we determined our sample size, and we explain all data exclusions (if any), all manipulations, and all measures in the study. All data reported in these experiments are available at: http://doi.org/10.17639/nott.7254 . Data were analyzed using IBM SPSS Statistics (Version 27) and JASP (2022) . This study’s design and its analysis were not preregistered. The task was programmed using Gorilla, and the materials are available upon request.

Data Exclusion

In all experiments, we conducted several checks to ensure data quality: participants who declared being color-blind were eliminated; during training, participants who failed to respond in more than 20% of the trials were removed from the analyses; at the end of training, only participants with higher average ratings to both cues presented alone and reinforced (A+ and V+) compared to the nonreinforced cues (B− and W−) were included in the analyses, otherwise they were not included. We asked participants if they had any knowledge of Chinese language; if they selected yes, they were removed from the analyses. Finally, at the end of the task, we asked participants about their subjective commitment during the task with the following question:

Well Done! The experiment is over. Just one last question. Did you give your full attention to the experimental task (as opposed to sometimes doing other things like using your smartphone) while stimuli were being presented? Please, answer honestly; this question has no impact on your payment. There are two options below “Yes” and “No.”

Participants who selected “No” were removed from the analyses and data of these participants were not replaced.

After applying these checks, 99 participants were considered for data analyses (32 in Group 1s, 32 in Group 3s, and 35 in Group 9s).

Test

We analyzed the averaged ratings of all cues during the test phase to assess differences between both categories. Critically, an overshadowing index was calculated as the difference between the control cue and the target cue in the reinforced cues (control − target) in each category. This is an index of competition whereby a value of 0 means the absence of cue interaction (neither overshadowing nor facilitation). In the reinforced cues, positive scores are indicative of cue competition or overshadowing whereas negative scores would be indicative of cue facilitation or potentiation. In the nonreinforced cues, positive scores are indicative of facilitation, whereas negative scores would be indicative of overshadowing. A 3 (group: 1s, 3s, 9s) × 2 (reinforcement: nonreinforced vs. reinforced) × 2 (category: color vs. symbol) mixed-ANOVA was conducted. In subsequent analyses, one-sample t tests were used to corroborate whether the index was different from zero or not. We provided BF 01 to test the reliability of the lack of overshadowing using the default Cauchy prior distribution ( JASP, 2022 ). As a general guide, we considered the Bayes factor above 3 as substantial evidence for the lack of differences ( Wagenmakers et al., 2011 ). The rejection criterion was set at 0.05 for all statistical tests. Partial eta-squared measures are presented for effect sizes, and their 90% confidence intervals were reported using the software available in Nelson (2016) . Cohen’s d was provided for the t comparisons. When the assumption of sphericity was violated, the Huynh–Feldt correction was applied in the corresponding tests of main effects or interactions.

Test Phase

Table 2 summarizes the ratings during the test phase in each group. All groups showed good discrimination between reinforced, nonreinforced, and partially reinforced cues. The lowest average rating for reinforced cues (A+, C+, V+, and X+) was above 50; in contrast, the largest rating for the nonreinforced cues (B−, D−, W−, and Y−) was below 45; finally, for the partially reinforced cues (E and Z) the ratings were somewhat in the middle. One relevant analysis results from the comparison of the two elements of the compound, that is, the target cue for symbol and color categories. In the case of reinforced cues, there was a main effect of category, with overall higher ratings to the color compared to the symbol, F (1, 96) = 9.43, p = .003, η p 2 = .09, 90% CI [0.02, 0.19] but modulated by group factor, F (2, 96) = 4.09, p = .020, η p 2 = .80, [0.24, 0.46]. Interestingly, only Group 1s showed differences between color and symbol categories, t (31) = 3.85, p < .001, this was not the case in Group 3s, t (31) = 1.43, p = .161 ( BF 01 = 2.09) nor in Group 9s, t (34) = 0.03, p = .974, ( BF 01 = 5.51). A similar pattern was observed with the nonreinforced cues, with differences in the case of Group 1s, t (31) = 5.43, p < .001, but not in the case of Group 3s, t (31) = 1.47, p = .151 ( BF 01 = 1.99), nor in Group 9s, t (34) = 0.65, p = .515, ( BF 01 = 4.51). Increasing the length of exposure to the cues resulted in the ratings between both elements of the compound being more similar, especially with 9s, with the Bayes factor above 3 for reinforced and nonreinforced cues. This is consistent with the notion that in the long exposure condition, the two cues (color and symbol) that form the compound have similar salience, overcoming the bias toward the color observed in the other two groups. However, this long exposure had little effect on the magnitude of overshadowing (see below).

Key to our goal was the overshadowing index. Figure 3 illustrates the overshadowing index in both categories (colors and symbols) for nonreinforced cues (left side of each figure) and for the reinforced cues (right side of each figure). For the nonreinforced cues, values below 0 represent that the target cue received higher ratings than the control cue, this suggests cue competition or overshadowing. In the case of reinforced cues, values above 0 represent competition between cues, that is, the target cue trained in compound received lower ratings than the respective control cue trained alone in each category. In both scenarios, a value of 0 means that the target and the control cue received similar ratings.

Figure 3 suggests an overall overshadowing effect. In general, pooling data across groups confirmed such a tendency. One-sample t tests revealed a significant overshadowing for nonreinforced cues, t (98) = 3.49, p = .001, d = 0.35 ( M color = −11.18, SD color = 31.83), and t (98) = 3.80, p < .001, d = 0.38 ( M symbol = −12.48, SD symbol = 32.62) for color and symbol, respectively; in the case of reinforced cues, overshadowing was also reliable in both categories: color, t (98) = 4.37, p < .001, d = 0.44 ( M color = 14.53, SD color = 33.08), and symbol, t (98) = 4.33, p < .001, d = 0.44 ( M symbol = 16.33, SD symbol = 37.50).

However, Figure 3 (panels A and B) suggests subtle differences between groups. A 3 (group: 1s, 3s, 9s) × 2 (reinforcement: reinforced vs. nonreinforced) × 2 (category: color vs. symbol) mixed-ANOVA revealed a main effect of reinforcement, F (1, 96) = 54.27, p < .001, η p 2 = .36, 90% CI [0.24, 0.46], and a triple interaction, F (2, 96) = 6.36, p = .003, η p 2 = .12, [0.03, 0.21]. Subsequent analyses explored the overshadowing effect in each group and level of reinforcement.

We compared the overshadowing index in each category with 0. One-sample t tests revealed than in Group 1s, there was no overshadowing considering nonreinforced cues in the color category, t (31) = 0.10, p = .921, d = 0.02, ( BF 01 = 5.27) but there was in the symbol category, t (31) = 4.25, p < .001, d = 0.75. The same pattern was observed in the reinforced cues, with no overshadowing in the color, t (31) = 1.45, p = .156, d = 0.26, ( BF 01 = 2.04), but reliable overshadowing in the symbol, t (31) = 2.85, p = .008, d = 0.50. Overshadowing was present in the symbol category, notably such competition was mitigated in the most salient category, the color.

In the case of Group 3s, in the nonreinforced cues, there was no overshadowing in neither category, t (31) = 1.56, p = .128, d = 0.28, ( BF 01 = 1.76) and t (31) = 1.19, p = .243, d = 0.21, ( BF 01 = 2.77) for color and symbol categories, respectively. In the case of reinforced cues, we observed overshadowing in the color, t (31) = 3.13, p = .004, d = 0.55, but not in the symbol, t (31) = 1.66, p = .106, d = 0.11, ( BF 01 = 1.54).

Finally, in the case of Group 9s for nonreinforced cues, overshadowing was reliable in the color category, t (34) = 4.98, p < .001, d = 0.84, but not in the symbol category, t (34) = 1.41, p = .168, d = 0.24, (BF 01 = 2.23). Moreover, for reinforced cues we observed overshadowing in both categories, t (34) = 3.31, p = .002, d = 0.56 and t (34) = 2.90, p = .007, d = 0.49, for color and symbol, respectively.