Variability and abstraction in evaluative conditioning: Consequences for the generalization of likes and dislikes☆
Introduction
Attitudes and preferences are important determinants of human behavior, guiding social decision-making in various situations such as hiring novel employees (Von Helversen, Herzog, & Rieskamp, 2014) or choosing a candidate to vote for (Galdi, Arcuri, & Gawronski, 2008; I. C. Lee et al., 2016). Humans often rely on likes and dislikes towards similar individuals, objects, and situations when making judgments and decisions on novel ones. At the same time, the generalization of attitudes can also have negative side effects. For instance, evaluations that are generalized from an individual to a social group can lead to prejudice and discrimination (Gilmour, 2015; Le Pelley et al., 2010). Considering the far-reaching consequences of attitude generalization, it is not surprising that the topic has received much research interest (e.g., Glaser & Kuchenbrandt, 2017; Högden, Stahl, & Unkelbach, 2020; Hütter, Kutzner, & Fiedler, 2014; Luck, Patterson, & Lipp, 2020; Von Helversen et al., 2014; Walther, 2002).
While much of this work has sought to understand how attitudes are generalized at the judgment stage (e.g., Högden et al., 2020; Von Helversen et al., 2014), little is known about the learning conditions that promote or diminish the generalization of attitudes. To address this question, one must consider (a) that different learning experiences can result in different cognitive representations of attitudes, and (b) that evaluations of familiar and novel attitude objects might depend in central ways on how attitudes are represented in memory (Hütter, 2022; Hütter & Rothermund, 2020).
Drawing on findings in other domains of learning research (e.g., Christie & Gentner, 2010), the present work focuses on the variability of training input as one means of generalization. Exposing learners to variable inputs has been described as an effective way of improving generalization in learning (Apfelbaum & McMurray, 2011; Estes & Burke, 1953; Hahn, Bailey, & Elvin, 2005; Raviv, Lupyan, & Green, 2022). We propose that the variability of stimulus objects encountered at attitude acquisition influences how attitudes are represented in memory, with consequences for the generalization of likes and dislikes. By taking on this cognitive-ecological perspective on the generalization of attitudes (Fiedler, 2014), we ascribe environmental conditions (here, stimulus variability in the environment) a key role in explaining evaluative learning and generalization. Understanding how environmental conditions relate to the generalization of likes and dislikes can contribute to our understanding of the acquisition of prejudice and stereotypes (Park & Hastie, 1987), and has implications for the design of interventions targeting attitude change (e.g., interventions to induce negative evaluations towards smoking; Măgurean, Constantin and Sava, 2016; or negative evaluations of unhealthy foods; Masterton, Hardman, Halford, & Jones, 2021; Bui & Fazio, 2016).
The relation between variability and generalization was documented in various domains of learning research, proposing that generalization is positively influenced by variability in training input (Apfelbaum & McMurray, 2011; Estes & Burke, 1953; Hahn et al., 2005; Raviv et al., 2022). For example, in category learning Posner and Keele (1968) reported an increase in the generalization of category knowledge after participants were exposed to variable rather than invariable training sets. In problem solving, the variability of worked examples increased the transfer of acquired skills to novel problems (Paas & Van Merriënboer, 1994). In concept learning, infants generalized a novel sound presented with animal categories to unknown category exemplars only after they experienced multiple (vs. single) animals per category (Vukatana, Graham, Curtin, & Zepeda, 2015). Similar outcomes were reported in research on language acquisition (e.g., speaker variability; Rost & McMurray, 2009), and inductive reasoning (e.g., premise diversity; Osherson, Smith, & Lopez, 1990). The studies highlight the relationship between variability of training exemplars and generalization at test. Importantly, because manipulations of variability produced similar results across domains of learning research, the underlying principles seem to be comparable (Raviv et al., 2022). Various accounts exist that attempt to explain the relation between variability and generalization.
One account suggests that variability fosters the formation of abstract representations during learning and thereby increases generalization (Apfelbaum & McMurray, 2011). Abstraction, in general, refers to the “process of identifying a set of invariant characteristics of a thing” (Burgoon, Henderson, & Markman, 2013; p. 502). Thus, abstract representations retain only those features that are relevant for a learning outcome, while irrelevant ones are ignored (see also Ramscar, Yarlett, Dye, Denny, & Thorpe, 2010; Reed, 2016). For example, a representation of several individuals in terms of their social group membership can be seen as abstract, as the representation highlights the common characteristics across individuals (e.g., fans of a soccer club wearing club merchandise). Because variability in training stimuli emphasizes invariant characteristics across training exemplars, it can facilitate abstraction. For example, variable training sets in reward learning help learners to identify the cues that are most predictive of a reward across instances. Later at test, learners can predict rewards based on the presence or absence of the cues in novel instances (Ramscar et al., 2010). At the same time, the formation of abstract, simplified representations has the drawback of diminishing memory for specific details. For example, abstract representations seem to make it harder for learners to distinguish between seen and unseen exemplars (Bowman & Zeithamova, 2020; Garagnani, Kirilina, & Pulvermüller, 2021; Hahn et al., 2005; Tussing & Greene, 1999).
One way to determine the relevance of features of exemplars is via cue competition. Cue competition, in general, describes the process by which cues compete for relevance in prediction of a particular outcome (Hoppe, Hendriks, Ramscar, & van Rij, 2022; Miller, Barnet, & Grahame, 1995; Ramscar et al., 2010; Rescorla, 1968; Siegel & Allan, 1996). Positive weights are formed for cues that produce little or no error for an outcome, while negative weights are acquired for cues that result in prediction errors (Ramscar, 2021). The overarching function of cue competition is that of reducing prediction errors, and hence uncertainty (Hohwy, 2020; Kiefer & Hohwy, 2019; Rescorla, 1968). Learning from variable stimuli allows cues to compete for relevance, which results in the cues that most reliably predict outcomes being emphasized. Put differently, variability improves the discrimination between cues in stimuli. By contrast, learning from stimuli that lack a rich cue structure hinders cue competition and thus also learning to discriminate between cues (Ramscar et al., 2010). Accordingly, this perspective explains increased generalization with higher variability in training via the formation of more abstract representations, with cue competition as the underlying principle.
An alternative to this account suggests that variability increases generalization via the number of exemplars that represent a concept. With an increasing number and diversity of training examples, the likelihood that a new stimulus resembles a known one increases as well (Bowman & Zeithamova, 2020; Hahn et al., 2005; Homa, Sterling, & Trepel, 1981; Nosofsky, 1988, Nosofsky, 2011). For example, diverse training stimuli in category learning offer learners the chance to draw broad inferences, because the training stimuli demonstrate the scope of the category (Homa et al., 1981; Nosofsky, 1988, Nosofsky, 2011). Importantly, this broadness account proposes that variability increases the number and diversity of representations making up a concept, but not their abstractness. As a consequence, and in contrast to the abstraction account, memory for specific details of training items should not be affected by variability (Bowman & Zeithamova, 2020).
To summarize, both accounts try to explain how variability affects generalization by specifying the way knowledge is stored in memory, either as abstract entities or as multiple concrete knowledge representations.
In the domain of attitude acquisition, the variability of attitude objects encountered during learning could also constitute a central determinant for the generalization of likes and dislikes. As an example, the co-occurrence of different female faces with positive images might trigger the acquisition of an association between a cue (“female”) and valence (“positive”), fostering generalization at test (Hütter et al., 2014). To our knowledge, no prior research exists that investigated the link between variability and generalization in attitude acquisition directly.
The present research aims to fill in this gap. We employ evaluative conditioning (EC) as an experimental procedure to study the acquisition and generalization of attitudes (EC; De Houwer, Thomas, & Baeyens, 2001; Hütter & Fiedler, 2016) and use recognition memory measures and evaluations of stimulus components to distinguish between the abstraction versus broadness accounts. In EC procedures, conditioned stimuli (CSs) that are often neutral in valence occur in spatiotemporal contiguity with unconditioned stimuli (USs) of negative or positive valence. The evaluations of the CSs typically change in the direction of the US valence, which is also referred to as the EC effect (for reviews, see De Houwer et al., 2001; Hofmann, De Houwer, Perugini, Baeyens, & Crombez, 2010; Walther, Nagengast, & Trasselli, 2005). Importantly, attitudes acquired via EC can generalize to novel stimuli never seen during learning. For example, conditioning of a specific image of a person changes evaluations of modified displays of the person (e.g., the same person photographed from a different angle; Hütter & Tigges, 2019). Moreover, using category exemplars as CSs subsequently changed evaluations of similar stimuli and the whole stimulus category (Glaser & Kuchenbrandt, 2017; Jurchiş, Costea, Dienes, Miclea, & Opre, 2020; Luck et al., 2020). While little is known about procedural aspects that might facilitate these generalization effects, the variability of attitude objects (CSs) encountered during conditioning might be a potential moderator.
In accordance with the “abstraction” account, high variability in CSs may result in abstract representations of CSs that only contain the cues that are most predictive of US valence across CSs. As a consequence, generalization towards novel instances increases. For example, imagine two distinct learning scenarios that vary in the variability of CSs. As displayed in Fig. 1, the first scenario (upper panel, “invariable CSs”) entails only a single CS (CS1) that consists of two components (Cue 1 and Cue 2). This CS is repeatedly paired with a US of the same valence (e.g., a positive image, US+). In this learning environment, the resulting representation constitutes a link between the concrete CS and the US valence (CS1-US+ associations).
The second scenario (lower panel, “variable CSs”) consists of CSs that overlap in one component (Cue 1), but vary in their second component (Cue 2, Cue 3, or Cue 4). All CSs would again be paired with USs of the same valence (e.g., positive valence, US+). In this condition, only one CS cue predicts US valence across stimuli (i.e., Cue 1). As a consequence, an abstract representation might form that entails the most predictive cue (i.e., Cue 1), while disregarding less predictive ones (i.e., Cues 2, 3, 4).1 In other words, the abstraction account predicts the formation of a link between the fixed cue and US valence (Cue1-US+ associations), at the cost of unique CS components.
Importantly, the two learning scenarios depicted in Fig. 1 should have consequences for the evaluation of novel, generalization stimuli (GS). Considering that generalization is generally driven by the perceptual overlap between a knowledge representation and a novel stimulus (Shepard, 1987), it becomes evident that the perceptual overlap is higher after the variable than the invariable conditioning procedure. The perceptual overlap between a GS that consists of the familiar Cue 1, and a novel Cue 5, and an abstract representation (containing only Cue 1) amounts to 100%, because Cue 1 is present both in the representation and in the generalization stimulus. On the other hand, the overlap between a specific CS representation (containing both familiar Cues 1 and 2) and the novel stimulus amounts to only 50% – because Cue 1 is present in the GS, but not Cue 2.2 Thus, generalization should be stronger in the variable (vs. invariable) condition.
Another possibility is that the various CSs encountered in the variable condition provide a broader basis for generalization due to their higher number, rather than abstractness. This would imply that learners represent all of the features of the individual CSs in the variable condition as well. Testing learners' memory for details, and their evaluations of single CS components provides one way to distinguish between the two accounts. First, learners should have greater difficulties distinguishing between old and new stimuli in the variable than the invariable condition if variability fosters abstraction in the representation of CSs. Second, evaluative judgments of the most predictive cues should be more extreme than evaluations of less predictive ones, if the resulting representation entails only the most predictive cue (e.g., Cue 1) rather than the CSs presented during learning. If these patterns are not observed, it would follow that the broadness account provides a better explanation for an increase in generalization than the abstraction account.
In this article, we report three studies in detail that manipulated the variability of CSs included in an EC procedure. CSs resembled Chinese characters that could be grouped into four categories by one common component. In the invariable condition, one item per category served as CSs. In the variable condition, multiple items per category were employed as CSs. Novel characters from the categories served as generalization stimuli (GSs). Generalization was expected to be more pronounced in the variable compared to the invariable condition. All experiments included both a direct (visual rating scales) and indirect measure (affect misattribution procedure; Payne, Cheng, Govorun, & Stewart, 2005) of attitudes. Because indirect measures infer attitudes from performance on a behavioral measure, they are less prone to demand effects and social desirability biases in responding.
We also included two additional measures to test the content of acquired representations of CSs. First, we included a variant of the Deese/Roediger-McDermott paradigm (DRM; Roediger & McDermott, 1995) to test participants' recognition memory performance. Participants were expected to make more recognition errors following an EC procedure with variable rather than invariable CSs. Second, we included evaluations of individual CS cues to test whether predictive cues are evaluated more extremely than less predictive ones for variable CSs. Evaluations should not differ between the two cues in the invariable condition.
The first experiment tested generalization alone, and Experiments 2 and 3 included the recognition memory measure and evaluations of individual CS components. Further, Experiment 2 controlled for the number of CSs used at test, and Experiment 3 held the total number of CSs included in the learning phase constant across the two learning conditions. We conducted one additional experiment that employed a similar experimental procedure as Experiment 2, but presented the dependent measure in a different sequence. We report this additional experiment in the Supplementary Material and as part of internal meta-analyses that integrate the findings from all experiments.
For all experiments, we report how we determined sample sizes, all data exclusions and all manipulations and measures employed. Pre-registrations (for Experiments 2 and 3), data files, analysis scripts and stimulus material are publicly available on OSF via https://osf.io/tafy9/?view_only=1ef497132b5d456a8f5ec940911bfca9. The studies were approved by the ethics committee for psychological research at the authors' institution.
Section snippets
Experiment 1
The first experiment sought to investigate our initial hypothesis that high variability in CSs during learning increases the generalization of likes and dislikes. That is, we tested whether the presentation of variable CSs increases the generalization of evaluations towards novel stimuli, relative to a condition that presents one specific CS repeatedly. The first experiment was not pre-registered.
Experiment 2
Experiment 2 was conducted to examine whether the results of Experiment 1 replicate when the confound noted above was controlled for. In Experiment 2, the number of CSs that were evaluated after the learning phase was held constant across CS variability conditions. The experiment also included two additional measures to test whether participants formed more abstract representations of CSs in the variable condition: First, participants completed a recognition memory task immediately after the
Experiment 3
In all of the previous experiments, the number of CSs per category that were included in the conditioning phase was confounded with the total number of CSs. That is, participants in the variable condition saw 20 CSs together with the USs, while participants in the invariable condition saw only 4 CSs. Thus, a potential alternative explanation for the results of the previous experiments might posit that the total number of CSs rather than the number of CSs per category was the crucial determinant
Internal meta-analysis
To assess the robustness of our findings in light of the varying sensitivity of the experiments to detect an effect, we conducted a maximum-likelihood random-effects meta-analysis using the R package metafor (Viechtbauer, 2010). The parameter coefficient for the three-way interaction of US valence, stimulus type and CS variability for direct evaluative ratings in Experiment 1 to 3, plus the additional experiment reported in the supplement (“Study 2S”) was significant, B = 19.91, 95%CI = [7.09,
General discussion
Previously acquired attitudes are often generalized to make judgments and decisions about newly encountered individuals, objects, and situations. Previous research has studied the principles of generalization at the judgment stage. However, until now little has been known about the learning conditions that facilitate or inhibit generalization, even though this question is highly relevant from both a theoretical and an applied perspective. Other domains of learning research have identified
Conclusion
The present findings give insight into the mechanisms underlying generalization in EC by providing a cognitive-ecological perspective on generalization effects. Theoretically, our results highlight the relevance of variability in generalization and propose that the learning principle of cue competition serves to specify how CSs come to be represented in memory. Practically, the present results offer insights into learning conditions that lay the ground for the formation of intergroup biases and
Open practices
All data files, analyses scripts, and the experimental material (Chinese characters, syllables used in the AMP) are publicly available via https://osf.io/tafy9/?view_only=1ef497132b5d456a8f5ec940911bfca9. Link to the pre-registration of Experiment 2: https://osf.io/g7nkw/?view_only=d7180cb703c343279ccc8fcdedd51780, Experiment 3: https://osf.io/dhf2s/?view_only=ae98a2a9d47341edac0c716bc77c7051.
Author notes
Kathrin Reichmann, Mandy Hütter, Barbara Kaup, Michael Ramscar, Department of Psychology, Eberhard Karls Universität Tübingen, Germany.
The present research was supported by a grant from the Deutsche Forschungsgemeinschaft (381713393) awarded to Mandy Hütter, Barbara Kaup, and Michael Ramscar as part of the research group Modal and Amodal Cognition (RU 2718), and a Heisenberg grant awarded to Mandy Hütter (HU 1978/7-1).
Declaration of Competing Interest
We have no conflict of interest to disclose.
References (89)
- et al.
Cue competition in evaluative conditioning as a function of the learning process
Acta Psychologica
(2015) - et al.
How variability shapes learning and generalization
Trends in Cognitive Sciences
(2022) - et al.
Differential effects of repetition on true and false recognition
Journal of Memory and Language
(1999) - et al.
Attitudes from mere co-occurrences are guided by differentiation
Journal of Personality and Social Psychology
(2020) - et al.
Using variability to guide dimensional weighting: Associative mechanisms in early word learning
Cognitive Science
(2011) - et al.
Fitting linear mixed-effects models using lme4
Journal of Statistical Software
(2015) - et al.
Evaluative conditioning is insensitive to blocking
Psychologica Belgica
(2009) - et al.
Training set coherence and set size effects on concept generalization and recognition
Journal of Experimental Psychology. Learning, Memory, and Cognition
(2020) - et al.
Generalization of evaluative conditioning toward foods: Increasing sensitivity to health in eating intentions
Health Psychology : Official Journal of the Division of Health Psychology, American Psychological Association
(2016) - et al.
There are many ways to see the Forest for the trees: A tour guide for abstraction
Perspectives on Psychological Science
(2013)
Sequential priming measures of implicit social cognition: A Meta-analysis of associations with behavior and explicit attitudes
Personality and Social Psychology Review
Where hypotheses come from: Learning new relations by structural alignment
Journal of Cognition and Development
Effects of talker variability on perceptual learning of dialects
Language and Speech
Implicit attitude measures: Consistency, stability, and convergent validity
Psychological Science
Associative learning of likes and dislikes: A review of 25 years of research on human evaluative conditioning
Psychological Bulletin
jsPsych: A JavaScript library for creating behavioral experiments in a web browser
Behavior Research Methods
Evaluative conditioning with foods as CSs and body shape as USs: No evidence for sex differences, extinction, or overshadowing
Cognition and Emotion
A theory of stimulus variability in learning
Psychological Review
Implicit measures in social cognition research: Their meaning and use
Annual Review of Psychology
From intrapsychic to ecological theories in social psychology: Outlines of a functional theory approach
European Journal of Social Psychology
Interventions designed to reduce implicit prejudices and implicit stereotypes in real world contexts: A systematic review. BMC
Psychology
Reducing Muslim/Arab stereotypes through evaluative conditioning
Journal of Social Psychology
Automatic mental associations predict future choices of undecided decision-makers
Science
Semantic grounding of novel spoken words in the primary visual cortex
Frontiers in Human Neuroscience
Analogy and abstraction
Topics in Cognitive Science
Formation of stereotypes
Behavioural Sciences Undergraduate Journal
Generalization effects in evaluative conditioning: Evidence for attitude transfer effects from single exemplars to social categories
Frontiers in Psychology
Simr: An R package for power analysis of generalised linear mixed models by simulation
Methods in Ecology and Evolution
Effects of category diversity on learning, memory, and generalization
Memory and Cognition
THINGS: A database of 1,854 object concepts and more than 26,000 naturalistic object images
PLoS One
Evaluative conditioning in humans: A Meta-analysis
Psychological Bulletin
Similarity-based and rule-based generalisation in the acquisition of attitudes via evaluative conditioning
Cognition and Emotion
New directions in predictive processing
Mind & Language
Evaluation of exemplar-based generalization and the abstraction of categorical information
Journal of Experimental Psychology: Learning, Memory, and Cognition
An exploration of error-driven learning in simple two-layer networks from a discriminative learning perspective
Behavior Research Methods
The education of attention as explanation of variability of practice effects : Learning the final approach phase in a flight simulator
Journal of Experimental Psychology: Human Perception and Performance
An integrative review of dual- and single-process accounts of evaluative conditioning
Nature Reviews Psychology
Editorial: Conceptual, theoretical, and methodological challenges in evaluative conditioning research
Social Cognition
What is learned from repeated pairings? On the scope and generalizability of evaluative conditioning
Journal of Experimental Psychology: General
Automatic processes in evaluative learning
Cognition and Emotion
On the external validity of evaluative conditioning: Evaluative responses generalize to modified instances of conditioned stimuli
Journal of Experimental Social Psychology
Evaluative conditioning of artificial grammars: Evidence that subjectively-unconscious structures bias affective evaluations of novel stimuli
Journal of Experimental Psychology: General
Representation in the prediction error minimization framework
The Routledge Companion to Philosophy of Psychology
Use of categorical and individuating information in making inferences about personality
Journal of Personality and Social Psychology
Cited by (4)
Modal and amodal cognition: an overarching principle in various domains of psychology
2024, Psychological ResearchUncovering the potential of evaluative conditioning in shaping attitudes toward sustainable product packaging
2024, Frontiers in PsychologyChoosing What You Like or Liking What You Chose? Sampling’s Impact on Evaluation and the Role of Idiosyncratic Reactions to Valent Stimuli
2024, Social Psychological and Personality Science
- ☆
This paper has been recommended for acceptance by Dr. Evava Pietri