Rule-based generalization of threat without similarity

Highlights

• Novel paradigm investigating rule-based threat generalization.

• Behavioral responses show successful threat conditioning and generalization.

• Evidence for abstract threat generalization without perceptual similarity.

• Flexible paradigm to investigate conceptual threat generalization.

Abstract

Threat generalization to novel instances is central to adaptive behavior. Most previous work has investigated threat generalization based on the perceptual similarity between past and novel stimuli. Few studies have explored generalization based on abstract, non-perceptual relations despite their importance for cognitive flexibility. In order to measure such rule-based generalization of threat without perceptual similarity, we developed a novel paradigm that prevents perceptual features from gaining predictive value. Our results demonstrate that participants responded according to the correct abstract rule and used it to successfully generalize their anticipatory behavioral threat responses (expectancy ratings, sudomotor nerve activity, and heart rate responses). Our results further show that participants flexibly adapted their responses to an unsignaled mid-session contingency reversal. We interpret our results in the context of other rule-based generalization tasks and argue that variations of our paradigm make possible a wide range of investigations into the conceptual aspects of threat generalization.

Introduction

Generalization is the expression of a previously learned behavior in response to a novel stimulus. In a dynamically changing world where conditions never truly replicate, generalization of behavioral responses to novel stimuli is necessary for any organism to establish beneficial interactions with its environment. Responding to a novel generalization stimulus (GS) in terms of a previously learned conditioned stimulus (CS) presupposes that GS and CS are equated with respect to their reinforcement contingencies even though the GS has not been reinforced. The behavioral equation of GS and CS with respect to their reinforcement contingencies, can be motivated by perceptual similarity (e.g., Pavlov, 1927, Spence, 1937, Guttman and Kalish, 1956, Honig and Urcuioli, 1981) or by an abstract pattern or rule (e.g., Gershman and Niv, 2015, Lange et al., 2017, Meulders et al., 2017, Lei et al., 2019, Lovibond et al., 2019, Maes et al., 2017, Vervoort et al., 2014).

Perceptual similarity can be defined as the spatial distance between the GS and CS on a continuously varying psychological dimension, such as, for example, size or wavelength (Shepard, 1987), or as the degree of overlap between the feature sets of the GS and the CS (Tversky, 1977; for a more general Bayesian account, see Tenenbaum and Griffiths, 2001). Experiments have shown that behavioral and physiological responses are sensitive to the degree of similarity and generally decrease as the GS increasingly differs perceptually from the CS, thus forming a generalization gradient (Spence, 1937, Lissek et al., 2008).

In contrast to similarity-based generalization, rule-based generalization requires the organism to infer an abstract pattern that specifies when a GS should be behaviorally equated with a CS. In a paradigmatic study on rule-based generalization, Shanks and Darby (1998) showed that humans who have been trained on negative (A+, B+, AB$-$) and positive (C$-$, D$-$, CD+) association patterns, generalized the learned rule of opposites (reinforcement for combined stimuli is opposite to singular stimuli) to novel stimulus patterns of the same type (E+, F+, EF?, correctly predicting EF$-$; GH+, G?, H? correctly predicting G$-$ and H$-$).

The main difference between similarity and rule-based generalization is that the latter requires some degree of abstraction from the identity of the stimulus and its concrete perceptual features, that is, inductive inference (Dunsmoor and Murphy, 2015). The abstraction required for rule-based generalization in the patterning task makes it possible to compare similarity-based and rule-based generalization (Shanks and Darby, 1998). However, the patterning task is focused exclusively on the rule of opposites, which might be a particular type of abstraction that requires other specific cognitive abilities, such as working memory or attention (Maes et al., 2015, Maes et al., 2017).

A simpler task to investigate rule-based generalization is based on the same/different distinction. In the match-to-sample (MTS) task, participants are first presented with a sample stimulus and then have to select the target stimulus that matches the sample stimulus from a range of comparison stimuli. Generalization in MTS is constituted by the selection of a target stimulus based on a previously reinforced matching relation between sample and target (e.g., color of elemental stimuli, or visual variance of compound stimuli; Urcuioli and Nevin, 1975). In its simpler form, MTS is problematic because successful performance in MTS generalization can be achieved through perceptual, that is, similarity-based, generalization using stimulus features such as visual entropy to assess visual variance (Young et al., 1997).

To address this issue with simple MTS tasks and to investigate rule-based generalization without similarity beyond the rule of opposites, we designed a novel paradigm that uses a relative – and therefore minimally abstract – property of compound stimuli to define sameness. The basic structure of our paradigm is as follows: on each trial, participants are consecutively presented with four different images, which together constitute a complex relational stimulus that acts as the CS. The first three images are all instances of the same semantic category, for example, three images of islands. The fourth image is either another member of that category (e.g., another image of an island) or not. The rule in this experiment is defined as a mapping from the relative perceptual properties of the four images to an abstract binary property of the CS (same/different). Reinforcement is contingent upon this abstract matching property of the CS, that is, depending on the CS-US contingency assignment, the CS is reinforced if the fourth stimulus is the same (CS+) but not otherwise (CS$-$). Importantly, all stimuli used in the experiment are trial-unique and each stimulus is presented only once during the experiment.

We argue that this paradigm requires rule-based rather than similarity-based generalization because the GS-CS relation is dependent on the abstract matching property but not any perceptual features of the stimuli. Perceptual similarity is only relevant to derive the matching property but is otherwise ‘encapsulated’ within each trial. The transfer of a previously learned CS-US association onto a novel GS is independent of the perceptual similarity between the GS and any other CS. There are no perceptual properties, for example, associated with islands, that will allow the learner to generalize to novel stimuli, which depict, for example, airports.

One feature of abstraction is that it affords a high degree of flexibility. Commonly, flexibility is experimentally assessed using reversal tasks, where reinforcement contingencies are reversed after a certain number of trials and whatever stimulus acted as CS+ now becomes the CS$-$ and vice versa (Clark et al., 2004, Izquierdo et al., 2017). Rules, such as the one used in our paradigm, afford flexibility because the relation between the matching property of the CS and the US can simply be reversed upon encountering new evidence without the need to learn a novel abstraction. To test whether rule-based generalization in our paradigm is associated with behavioral flexibility, participants undergo a mid-session reversal of reinforcement contingencies.

As behavioral measures of learning, we recorded skin conductance responses (modeled as sudomotor nerve activity, SNA), stimulus-related changes in heart rate (heart rate responses, HRR), and online US expectancy ratings. Our main hypothesis is that participants will demonstrate differential, rule-based generalization across all three measures, that is, stronger responses to CS+ than CS$-$ irrespective of experimental phase. We further predicted that participants would flexibly adapt their behavioral responses after the contingency reversal, that is, differential responding to CS+ and CS$-$ would be maintained post contingency reversal.