The adaptive mechanism of fear generalization prevents the encounter with unknown threats by extending previous learning to new cues (Dunsmoor and Paz, 2015). Thus, fear responses are elicited not only by cues (conditioned stimulus, CS+) predicting threat (i.e., the unconditioned stimulus, US; LeDoux and Pine, 2016) but also by stimuli, which were never associated with the US but share physical or semantical properties with the threat-associated cues (Lissek et al., 2008; Dunsmoor and Paz, 2015; Dymond et al., 2015). Notably, this mechanism is exaggerated in anxiety patients compared to healthy controls, leading to over-generalization of fear (Lissek et al., 2010, 2014). Similarly, high trait-anxiety individuals, i.e., individuals at risk for anxiety disorders (Raymond et al., 2017), tend to over-generalize conditioned fear (Baumann et al., 2017; Stegmann et al., 2019; but see Torrents-Rodas et al., 2013; for a meta-analysis, see Sep et al., 2019).

Generalization of conditioned fear seems to be related to an incapacity in perceptually discriminating the relevant stimuli (Holt et al., 2014; Struyf et al., 2018; Zaman et al., 2019). For example, Holt et al. (2014) first assessed the participants’ ability to discriminate the to-be-conditioned face from other faces, i.e., they assessed the just noticeable differences (JND) between CS+ vs. CS− and their morphs (i.e., generalization stimuli, GSs). After differential conditioning, a generalization test revealed increased US expectancy for the GSs below and at JND, indicating generalization, but not for those above JND. Thus, generalization of conditioned fear did not occur to those GSs, which were reliably discriminated from the CS+. Accordingly, improving the discrimination of CS+ may be a promising approach to reduce fear generalization.

Following this idea, several kinds of discrimination training have been developed to prevent fear generalization (Ginat-Frolich et al., 2017, 2019; Lommen et al., 2017). Specifically, two studies applied a discrimination training of fear-irrelevant stimuli (Ginat-Frolich et al., 2017, 2019). Each discrimination training trial started with a 4-s presentation of one abstract shape (target) at the center of a screen. Then, the target plus a new similar shape were presented on the left and right of the screen, and the participants had to identify the target. This training group was compared to a non-discriminative control task group where the participants had to indicate on which side of the screen each shape appeared. As a result, the authors observed reduced fear generalization of US expectancy in the first compared to the latter group.

Another study applied a discrimination training with stimuli sharing one perceptual characteristic (i.e., color) with the CS+ (Lommen et al., 2017). The participants had to decide if two stimuli presented shortly after each other differed in size or color or not at all. The participants assigned to a control task had to indicate instead which of two words was related to each presented stimulus. In contrast, this discrimination training which occurred before fear acquisition did not reduce the fear generalization of US expectancy but of avoidance behavior, again compared to the non-discriminative control task. These studies suggest that discrimination trainings with either fear-irrelevant or partial fear-relevant stimuli reduce fear generalization. However, please note that these trainings occurred before (Lommen et al., 2017) or after (Ginat-Frolich et al., 2017, 2019) fear learning. Therefore, only the latter two studies reflect a treatment training, while the former mirrors a prevention training.

One limitation of the above-mentioned studies is that they have tested the effects of discrimination training before having examined the generalization gradient of their participants. However, anxiety disorder patients start treatments after fear acquisition and with a long history of over-generalization. Thus, we examined the fear-reducing effects of a discrimination training after both fear acquisition and a first demonstration of fear generalization. Our study was designed in order to reduce fear generalization, which occurred due to problems with perceptually discriminating CS+ from safe similar stimuli. As there is evidence that perceptual learning cannot be easily transferred to an unpracticed stimulus set (Furmanski and Engel, 2000), we examined whether discrimination trainings with fear-relevant vs. fear-irrelevant stimuli show higher efficacy. Additionally, we examined whether feedback on discrimination performance further enhances training efficiency. The latter expectation is based on previous studies demonstrating that positive feedback (Sasaki et al., 2009) on participants’ performance and, in particular, when associated with reward (Weil et al., 2010) is able to improve stimulus perception.

To examine these hypotheses, our participants first underwent differential fear conditioning followed by two generalization blocks separated by discrimination training. Importantly, participants either learned to discriminate fear-relevant (discrimination of CS+ from GSs and CS−) or fear-irrelevant (discrimination of a specific line width from other lines) stimuli either with or without performance feedback. We hypothesized that the fear-relevant training reduces fear generalization more effectively than the fear-irrelevant training, and that this effect is especially strong if feedback is applied.

This study was designed to test our hypothesis that the generalization of conditioned fear to perceptually similar stimuli can be reduced significantly better by a discrimination training with fear-relevant vs. fear-irrelevant stimuli. Additionally, we hypothesized that discrimination training effects profit by reinforcing feedback with greatest effects due to fear-relevant training with feedback. The experimental design followed the logic of a treatment, which normally starts in patients with ascertained fear generalization. Thus, the discrimination training occurred after fear acquisition and in between two generalization tests. In consequence, this is the first study allowing one to verify training effects by comparing the generalization indices pre- vs. post-training.

Analysis of acquisition data validated that successful fear conditioning as the threat cue (CS+) vs. the safety cue (CS−) triggered greater arousal, more negative valence, greater US expectancy, and larger SCRs. Furthermore, analyses of the CS+ vs. CS− responses during the generalization tests pre- and post-discrimination training (see Supplementary Material) indicate that these conditioned fear responses remained stable throughout the experiment. Overall, these results corroborate previous studies using similar paradigms (Lissek et al., 2008; Schiele et al., 2016) and indicate the successful acquisition of fear toward the CS+.

Analyses of discrimination training effects on fear generalization suggest, on the one hand, stronger effects for the fear-relevant vs. the fear-irrelevant discrimination training and, on the other hand, stronger effects for discrimination training with reinforcing feedback vs. without feedback, but with no interaction of effects. Thus, we conclude that a reduction of fear generalization is especially effective when the discrimination training is carried out with fear-relevant stimuli vs. fear-irrelevant stimuli or with reinforcing feedback vs. no feedback. However, we found no support for our hypothesis that a combination of training, i.e., training with fear-relevant stimuli and feedback, is especially effective in reducing fear generalization. Please note, however, that these findings are restricted to US expectancy ratings.

These results raise the following questions: Firstly, what contributes to the observed superior effect of a discrimination training with fear-relevant vs. fear-irrelevant stimuli on reducing fear generalization? Secondly, why is feedback during discrimination training effective in reducing fear generalization independently from the training stimuli? Thirdly, why are the observed discrimination training effects restricted to the US expectancy ratings?

Regarding the first question, the effect might be related to working memory (WM) processes, which are addressed during discrimination trainings as the participants have to keep in mind the first stimulus in order to compare it with the subsequent one (Lenaert et al., 2016). Considering that WM filters irrelevant from relevant information (Eysenck et al., 2007; Derakshan and Eysenck, 2009; Baddeley, 2012), it is very likely that a fear-relevant discrimination training improves the discrimination between relevant and irrelevant CS+ information, especially strong, and subsequently reduces the generalization effects. Through its relevance in attentional control (Eysenck et al., 2007; Derakshan and Eysenck, 2009), WM might facilitate perceptual learning (Amitay et al., 2014), possibly also thanks to inductive processes like the derivation of rules from WM (e.g., paying attention to visual features of stimuli; Livesey and McLaren, 2009). Rule learning processes might be further promoted by the fixed sequence of stimulus presentations during training (always beginning with CS+ or Line1), which could lead to an overestimation of perceptual learning (García-Pérez and Alcalá-Quintana, 2020). Secondly, the results go in line with previous reports that a discriminative training with fear-relevant stimuli is able to reduce fear generalization (Lommen et al., 2017), although this study trained the participants before fear acquisition in analogy to prevention, while we trained the participants after fear acquisition in analogy to therapy. Moreover, it goes in line with more general findings regarding the transfer of discrimination learning to a test phase, which was found to be more effective after training relevant compared to irrelevant stimuli (Furmanski and Engel, 2000). In our study, the fear-relevant and fear-irrelevant training groups both have improved discrimination performance during training (for more details, see Supplementary Material). Consequently, the more efficient fear reduction of the fear-relevant training group might be related to a more effective transfer to the post-training generalization test. However, we know from previous studies that discrimination training, also with fear-irrelevant stimuli, reduces subsequent generalization effects in general (Ginat-Frolich et al., 2017, 2019). This can explain why the effects of fear-relevance are present but rather small.

Regarding the second question, we argue that feedback affects very general processes unrelated to the processing and discrimination of the presented stimuli. Previous studies revealed that feedback improves performance in general (Sasaki et al., 2009; Weil et al., 2010) as well as mood (Westermann et al., 1996). Moreover, reward facilitates attentional processes toward the reward-associated cue (Chelazzi et al., 2013) and increases motivation (Fishbach et al., 2010). Therefore, we assume that the implemented reinforcing feedback improved our participants’ mood, attention, and motivation in general, which consequently improved their discrimination during the following generalization test. To verify this hypothesis, future studies might consider adding a control condition without discrimination training but a motivation condition, e.g., a task with comparable cognitive load and reinforcing feedback for correct responses.

Thirdly, we have to consider why the effects of the implemented discrimination trainings were restricted to US expectancy ratings. One possible explanation lies in the typology of the ratings. As previously suggested (Lonsdorf et al., 2017), US expectancy reflects cognitive learning processes, while valence and arousal ratings rather reflect affective learning processes. Therefore, new learning experiences can be seen more easily and earlier in US expectancy ratings. In line with this, the beneficial effects of our training were visible for US expectancy ratings. In contrast, extinction after evaluative conditioning, reflected in changes in valence, is hard to achieve (Vansteenwegen et al., 2006; Hofmann et al., 2010). This might also be true for discrimination learning, which could explain why we did not find training effects for the affective ratings. Moreover, the dissociation between affective and cognitive ratings suggests that, apart from perception, other (higher-order) processes (e.g., WM and inference rules) might have also determined fear generalization and its reduction in our paradigm. We suggest that future studies should explore whether and at what level an extensive discrimination training can improve discrimination as reflected in affective responses and, in consequence, reduce generalization.

In order to reveal discrimination training effects on the fear generalization of affective responses, future studies might consider assessing online ratings of valence and arousal during tests. We assessed the ratings at the end of every generalization block only and therefore could not determine discrimination training effects by comparing ratings from the end of generalization pre-training with the beginning of generalization post-training. Considering that online ratings might draw attention to US contingencies (Lonsdorf et al., 2017), we decided to assess intermittent ratings as several previous studies did (Haddad et al., 2012; Meulders et al., 2013; Holt et al., 2014; Roesmann et al., 2020).

Finally, we want to address the limitations of the current study. First, one limitation concerns the potential role of extinction and safety learning as additional mechanisms contributing to the higher efficacy of fear-relevant discrimination training. During the fear-relevant training, the subjects are exposed to the CS+ and the GSs without the US’s presence, likely triggering extinction and safety learning processes. We can assume that our measured effects do not include extinction processes for CS+ because USs were presented again during generalization testing, thereby very likely undoing the CS+ extinction learning (as indicated by stable conditioned fear response to CS+ both pre- and post-training; see Supplementary Material). However, we cannot rule out that lower responses to certain GSs (and in consequence less generalization as reflected in the generalization index) stem from safety learning rather than some effects of perceptual discrimination training. Therefore, future studies should consider examining an additional group that did not undergo discrimination training but was exposed to the same amount of stimuli to control for potential extinction and safety learning effects. Second, no effects of fear relevance or feedback could be confirmed for physiological arousal, i.e., SCRs, which limits the effectiveness of our “therapeutic” approach of discrimination training to the verbal level. However, previous studies also did not find training effects for physiological measures (Ginat-Frolich et al., 2017) or did not include physiological measures (Lommen et al., 2017). In our study, contrary to all ratings, SCRs post-training were not significantly influenced by the pre-training level, as shown by the absent main effect of covariate. This means that all trainings strongly influenced the electrodermal activity in a way that dissolved the pre-training generalization pattern. Conceivably, the main reason lies in the nature of a training. While the participants passively observed the stimuli during the generalization blocks, during the discrimination training they actively fulfilled a task. The active task might have overlaid specific training effects. Besides this, the discriminative physiological arousal found during the first acquisition block disappeared in the second acquisition block, which is presumably due to habituation and could explain the absence of training effects for SCR. Third, the sample sizes are rather small. It should be considered, however, that the study presented here is a proof-of-principle study to establish a therapeutical approach of fear reduction by discrimination trainings. Fourth, we examined healthy participants who show moderate fear generalization as previous studies have also reported (Lissek et al., 2008). Considering the over-generalization in highly anxious individuals and anxiety patients (Lissek et al., 2010, 2014; Laufer et al., 2016; Sep et al., 2019), future studies should investigate discrimination training in more clinic-relevant samples. Fifth, we have to acknowledge that, due to a programming error, we have no data on the discrimination training performance for the no-feedback groups. Therefore, we are unable to test whether and how feedback improved discrimination performance during training, and consequently, we cannot relate feedback effects on performance to later generalization effects.

In conclusion, this proof-of-principle study demonstrated the successful reduction of existing fear generalization by a fear-relevant discrimination training in healthy individuals at least for cognitive fear parameters, i.e., US expectancy. Moreover, we revealed that reinforcing feedback during discrimination training reduces the generalization of US expectancy, presumably via motivational mechanisms. Importantly, our study is the first to mimic a therapeutic approach as the discrimination training was performed after fear acquisition and a first demonstration of fear generalization. Accordingly, this is the first indication that pre-existing fear generalization can be reduced by means of a fear-relevant discrimination training or a reinforcing discrimination feedback, which in the future might be successfully used in patients with anxiety disorders.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Ethical Committee of the Medical Board of the University of Würzburg, Würzburg, Germany. The patients/participants provided their written informed consent to participate in this study.

KH, MA, KD, MR, JD, and PP contributed to conception and design of the study. KS was involved in the conception and design of the discrimination training. MS contributed to the design of the fear generalization paradigm. KH organized the database and performed the statistical analysis. KH, MA, and PP wrote the first draft of the manuscript and critically revised each section. All authors contributed to manuscript revision, read, and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.