The ability to recognize threatening stimuli clearly increases the chances of survival. Given that a known threat can take many forms, it is also adaptive to be cautious with other exemplars of the same semantic category that may predict a similar aversive outcome (Mineka, 1992). Stimulus generalization —a learning mechanism whereby conditioned responses extend to a range of stimuli resembling the original conditioned stimuli (Pavlov, 1927) – enables a fast response to novel potentially threatening stimuli. Yet, it can turn into maladaptive behavior when non-threatening stimuli or contexts are inappropriately treated as harmful. Here, we will refer to this phenomenon as ‘overgeneralization of fear’ or ‘maladaptive fear generalization.’ The way we use the term ‘fear’ here includes both physiological and behavioral responses to threat, as well as the subjective experience of fear (but see for an alternative use of the term ‘fear’ LeDoux, 2014).

Maladaptive fear generalization is a characteristic of anxiety disorders and post-traumatic stress disorder (Lissek et al., 2005, 2014; Mineka and Zinbarg, 2006; Kong et al., 2014; Bishop et al., 2015) and may even play a causal role in these disorders (Mathews and MacLeod, 2002; Wilson et al., 2006). For example, while in spider phobia fear responses to real spiders may be debilitating in itself, fear responses to stimuli that more or less resemble the object of fear (e.g., a piece of dust) may interfere most with daily functioning as phobic individuals find themselves in a permanent state of hypervigilance, avoiding many ‘safe’ situations (e.g., not eating tomatoes as their insides resemble the legs of a spider). Clarifying which processes enhance fear generalization will ultimately help to answer the fundamental question of why and how people differ in their disposition to develop maladaptive fears.

Based on decades of animal conditioning research that focused on the perceptual similarity and discriminability of threatening stimuli, it has been implicitly assumed that overgeneralization of fear is a perceptual deficit (Shepard, 1987). In line with this, examples from research in humans show that fearful individuals judge neutral faces as more negative (Richards et al., 2002; Bell et al., 2011), and that individuals with spider phobia more easily see a spider in pictures morphing from flowers to spiders (Kolassa et al., 2007). Only recently it became evident that fear generalization not solely depends on the physical properties of threatening stimuli but also on their conceptual properties (Dunsmoor et al., 2009, 2011, 2012; Soeter and Kindt, 2012, 2015; Kindt, 2014). This raises the question whether fear generalization observed for perceptual cues (such as when a piece of dust triggers a fear response) is in fact a perceptual process, or may instead emerge at a later stage of processing, when activated fear associations start guiding (biasing) the interpretation of a stimulus.

As the observed behavior does not reveal whether overgeneralization of fear already occurs during the initial perception and categorization of a stimulus, or emerges at a later stage, it is necessary to go beyond behavioral observations to study the (neural) processes that drive these behaviors. A number of studies have begun to elucidate brain processes involved in fear-related perceptual biases. These studies mainly found heightened sensory sensitivity to all external stimuli in high fearful individuals, expressed as enhanced early (100 ms) event-related potentials (ERPs; Kolassa et al., 2007, 2009; Frenkel and Bar-Haim, 2011; Weymar et al., 2013) and heightened responses in visual areas to phobogenic objects, often paralleled by heightened responses in the amygdala (Dilger et al., 2003; Straube et al., 2006; Alpers et al., 2009). These findings are in line with the commonly observed fear-related attentional bias (Bar-Haim et al., 2007), suggesting that fear facilitates afferent cortical processing in the human visual cortex when individuals search for potential threat. However, although heightened sensory sensitivity may explain a faster detection of a stimulus, it does not necessarily imply or explain a biased classification of that stimulus.

Studies on normal fear generalization have found that varying degrees of perceptual resemblance to a conditioned stimulus elicit graded responses (generalization curves) in the same neurocircuitry as is involved in the acquisition and expression of conditioned fear (i.e., insula, dorsal anterior cingulate cortex; Dymond et al., 2014), and in salience processing in general (e.g., the ventral tegmental area; Cha et al., 2014). While normally these graded responses inversely relate to activation in inhibitory brain systems, such as the hippocampus and the ventromedial prefrontal cortex, individuals with generalized anxiety seem specifically impaired in recruiting these systems, broadening the range of stimuli to which they respond with fear (Greenberg et al., 2013; Cha et al., 2014; Bishop et al., 2015).

Even though the aforementioned studies provided useful insights into the brain areas that are hyper- or hypoactive in anxiety disorders, they do not distinguish between a perceptual and a conceptual account of overgeneralization of fear in anxiety disorders. An increased neural response to a stimulus does not specify how that stimulus is categorized (e.g., high anxious individuals may exhibit heightened sensory sensitivity to virtually all stimuli, while only displaying a classification bias for a subset of stimuli). Univariate techniques therefore seem unsuited for addressing the question as to where in the cortical hierarchy ambiguous stimuli are originally marked as threatening.

In contrast, multi-voxel pattern analysis (MVPA) evaluates the information across groups of voxels, to characterize the unique neural representation of a stimulus within a certain brain region (Haxby et al., 2001). By training a classifier on neural patterns related to distinct stimulus classes one can classify patterns related to novel stimuli, providing a more sensitive way to assess the degree to which different stimuli or cognitive states are alike (Haynes and Rees, 2005; Kamitani and Tong, 2005; Norman et al., 2006; Kriegeskorte et al., 2008), or altered by fear (Li et al., 2008; Visser et al., 2011, 2013, 2015, 2016; Dunsmoor et al., 2014).

Here, we combined functional magnetic resonance imaging (fMRI) with an adapted version of the task used by Kolassa et al. (2007), to study overgeneralization of fear in individuals with low and high fear of spiders. Based on previous work, we predicted that individuals with high spider fear (HSF) would be more likely to classify ambiguous morphs as spiders than individuals with low spider fear (LSF). Furthermore, we examined in a data-driven manner whether overgeneralization of fear is associated with functional anomalies in regions traditionally associated with (1) early perception and object identification (Ungerleider and Haxby, 1994), which would support a perceptual account, and/or 2) regions involved in saliency (Etkin and Wager, 2007; Seeley et al., 2007; Hermans et al., 2011; Ipser et al., 2013), and higher cognitive processes (Miller and Cohen, 2001), which would support a more conceptual account. Of course, if a bias is already present in low-level areas it is likely to be present in higher-level areas as well, given that these areas respond to the information relayed from lower-level areas. If, however, a bias does not emerge until later in the processing stream, this could be an indication that the stimulus is initially correctly identified as non-threatening. Yet higher in the processing-stream certain features of the stimulus may still trigger semantic associations with the feared object, which in turn may evoke unpleasant feelings and biased decision making. Using support vector machine (SVM) classification in functional regions of interest we assessed at what point in the information-processing stream (i.e., ‘low-level’ visual areas, and/or ‘higher’ areas associated with the attribution of significance and decision making) this bias would become apparent.

The aim of the present study was to examine how fear influences the processing of ambiguous stimuli. We used SVM classification to explore whether overgeneralization of fear is associated with functional anomalies in regions traditionally associated with early perception and object identification, and/or with regions involved in saliency and higher cognitive processes. Identifying where in the information-processing stream fear affects ambiguity resolution could potentially indicate whether this bias is primarily perceptual, or conceptual in nature.

In line with previous findings (Kolassa et al., 2007), individuals with HSF were more likely to classify ambiguous morphs as spiders than individuals with LSF. Unexpectedly, neither average activation, nor SVM classification in 12 functional ROIs revealed a pattern that mirrored the behavioral effect. We even obtained some evidence that response patterns in visual association areas related to ambiguous stimuli were more likely to be classified as spiders when individuals were not afraid of spiders.

These findings seem to point to a methodological limitation of the current paradigm. Unlike most studies, which used either pictures of phobogenic material (Dilger et al., 2003) or schematic morphs (Kolassa et al., 2007), we alternately presented morphs and pictures of real spiders. In the SVM classification, we trained a classifier on the unambiguous pictures to classify the schematic morphs. This seemed the most ecologically valid approach, as we were interested in the degree to which the morphs would resemble ‘true’ spiders or flowers, not in the similarity between schematic morphs per se. However, by mixing the two types of stimuli, we may have unintentionally deflated the valence of the spider drawings, skewing the classification of these stimuli. Although these morphs could have been a valid representation of the fearful category if presented alone, the (anticipated) presence of stimuli that are substantially more arousing may have turned the morphs into relatively safe stimuli. For individuals without spider fear both classes of unambiguous stimuli were virtually neutral, so training a classifier on these stimuli showed a linear increase in the likelihood that an ambiguous morph is classified as a spider. In contrast, for individuals with HSF the classifier is trained on two very distinct categories: one neutral and one highly emotional. This is also evident from the fact that these unambiguous categories are better classified – thus more distinct – in high fearful individuals than in low fearful individuals. In this scenario, a relatively neutral morph is classified as a flower, probably not on the basis of perceptual or even conceptual similarity, but on the absence of a strong emotional response. Between-run classification, with ambiguous and unambiguous stimuli presented in separate runs, could partially solve this problem: without the continuous anticipation of the (terrifying) unambiguous spider pictures, the spider drawings could be semantically categorized as spiders, instead of merely being categorized as ‘relatively safe.’ Between-run classification also has the advantage of ensuring complete independence between train and test set, thereby preventing possible biases in pattern classification (Mumford et al., 2014). Yet, even in separate runs, the unambiguous spider pictures will undoubtedly elicit a much stronger emotional response than the morphs, which may still hamper a balanced classification of stimulus categories if the unambiguous stimuli were to be used for training. This exemplifies that it is methodologically challenging to classify emotional and neutral stimuli in groups that differ in their judgment of emotionality.

Even though the design may have skewed the classification of morphs in some areas of the brain, participants did eventually indicate that they had seen a spider. Given that every behavioral response must have a neural correlate, one would assume that it should be possible with SVM classification to detect an area that codes for this behavior. We did not find such an area. Again, this suggests that our study design may not have been optimal for detecting subtle effects. While overgeneralization of fear was observed at the level of the group, individual behavior consisted of three discrete responses per stimulus type, yielding a proportion of stimuli classified as flower or spider. Thus, individual behavior was not necessarily representative of the group average. Although we based our behavioral task on previous research (Kolassa et al., 2007), future research should explore the possibility of having participants evaluate ambiguous stimuli on a continuous scale, which seems a more sensitive approach for linking individual behavior to brain data. Further, it may be useful in future research to add a jitter between the presentation of a stimulus and presentation of the response screen, so that responses could be independently modeled as regressor of no interest. This would make it easier to separate decision-making processes from basic perceptual processes. Lastly, the number and spacing of stimuli used in this experiment were based on previous studies. In these studies pattern analysis was applied to distinguish face and house stimuli (Epstein and Kanwisher, 1998; Haxby et al., 2001) and to assess the effects of Pavlovian conditioning (Visser et al., 2011, 2013, 2015, 2016), which is a powerful manipulation. The present manipulation – gradually morphing a flower into a spider – was presumably subtler and therefore required a greater number of trials. Increasing the power and using different behavioral measures would likely yield stronger effects and would open up avenues for model-based searchlights, using (continuous) representational similarity analysis (Kriegeskorte et al., 2008; Kriegeskorte, 2011) instead of the rather coarse (dichotomous) classification analysis. Approaches like these could further elucidate how fear influences the perception, categorization and interpretation of ambiguous stimuli.

In sum, while we found behavioral evidence for over-generalization of fear in spider phobia, replicating previous findings (Kolassa et al., 2007), we were not able to identify a neural signature of this bias. The question as to where in the information-processing stream this bias emerges therefore remains a topic for further investigation.

RV, PH, and MK designed the research; RV and PH performed the research and analyzed the data; HS and RZ contributed analytic tools; RV and MK wrote the manuscript.

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.