People apply what they have learned from past experiences to similar situations, a phenomenon known as generalization. For example, if eating a particular food caused illness, a person will likely avoid foods that look or smell similar in the future. Generalization can be helpful because it allows people to decide how to act in new situations. But over-generalizing after a bad experience could lead an individual to fear benign scenarios. This may lead to unnecessary anxiety. It can also create a negative cycle where people avoid certain situations or objects, which prevents them from learning that they are safe.

Now, Norbury et al. show what happens in the brain when making decisions that involve generalization. In the experiments, volunteers were told seeing a particular flower design would lead to a painful electric shock, unless they pushed a button to ‘avoid’ that image. Individuals completed this task in a magnetic resonance imaging machine so Norbury et al. could observe their brain activity while they completed the task. A second group of individuals were asked to complete a similar task online, but instead of being shocked they lost money if they failed to hit a key when they saw the ‘dangerous’ flower. The online participants also filled out a survey about their experience of various psychological symptoms.

Norbury et al. used computer modeling to reconstruct how people decided whether or not to avoid images that looked similar to the harm-associated images but were in fact safe (did not lead to pain or losing money). The experiments showed that different parts of the brain were involved in different parts of the generalization process. Areas of the brain that interpret vision, fear, and safety played distinct roles. People who generalized more from harmful outcomes were more likely to report feeling anxious and having intrusive negative thoughts in their everyday lives. A better understanding of the brain processes that cause these symptoms in different situations might help scientists develop better treatments for conditions like anxiety in the future.

During aversive decision-making, generalization allows application of direct experience with a limited subset of dangerous real-world stimuli to a much larger set of potentially related stimuli. For example, if eating a particular foraged fruit has led to food poisoning in the past, it may be adaptive to avoid similar-appearing fruit in the future. As an evolutionarily well-conserved process, generalization enables safe and efficient navigation of a complex and multidimensional world (Sutton and Barto, 1998; Ghirlanda and Enquist, 2003). However, over-generalization, resulting in inappropriate avoidance of safe stimuli, actions or contexts, has been suggested as a possible pathological mechanism in a range of psychological disorders including anxiety, chronic pain, and depression (Duits et al., 2015; Dymond et al., 2015; Vlaeyen and Linton, 2012; Harvie et al., 2017; Pearson et al., 2015).

Previous work on aversive generalization has focused on predicting punishments in passive (Pavlovian) designs. Such studies have revealed evidence of heightened subjective, physiological and neural responses to stimuli that bear perceptual similarity to learned exemplars (Dymond et al., 2015). However, the extent to which these observations extend to a decision-making context − that is whether or not to make an avoidance response in the face of certain stimuli, allowing us to exert control over experience of aversive outcomes − is unclear. Although Pavlovian processes can influence avoidance learning, the latter involves acquisition of a fundamentally distinct set of values relating to actions themselves. This is a clinically important distinction, as theories of many psychological disorders relate specifically to excessive avoidant behaviour over and above subjective fear (Krypotos et al., 2015) − for example, by reducing opportunities for extinction of inappropriate fear or allowing unnecessary avoidance to transfer to habit-based control (Arnaudova et al., 2017; LeDoux et al., 2017; Gillan et al., 2014).

There are a number of potential mechanisms by which avoidance generalization could be implemented by the brain. As emphasised in some accounts, perceptual uncertainty in stimulus identity alone can effectively yield generalization. Although there is debate about how well discriminative ability is controlled for in many generalization experiments (Struyf et al., 2015), there is good evidence that experience with aversive outcomes alters the representation of predictive stimuli in primary sensory cortices (Weinberger, 2007; Sasaki et al., 2010; Wigestrand et al., 2017), and that this may result in changes to absolute stimulus discriminability (Resnik et al., 2011; Laufer and Paz, 2012; Aizenberg and Geffen, 2013). On the other hand, generalization may also occur at the level of value representations, by the transfer of acquired value to similar, but discriminable cues during learning. In the Pavlovian case, several well-established behavioural phenomena implicate value-related processes at play in generalization across species (Hanson, 1959; Schechtman et al., 2010). That both perceptual and value processes might operate in parallel may explain why recent neuroimaging studies have highlighted different brain areas (e.g. limbic cortex vs primary sensory regions) as being key to Pavlovian aversive generalization in humans (Onat and Büchel, 2015; Laufer et al., 2016).

A further important factor in the control of avoidance learning is reinforcement by neutral (or ‘safety’) states, that signal omission of punishment. It is likely that generalization over these states can also influence behaviour: for example in the Pavlovian case, evidence for this is seen in ‘peak-shift’ effects, whereby the presence of a perceptually similar safety cue appears to inhibit response to nearby aversive cues (Hanson, 1959). It is therefore possible that under-generalization of safety cues, as opposed to over-generalization of aversive cues, might be a contributing factor to susceptibility to disorders such as generalized anxiety in humans (Grupe and Nitschke, 2013).

Here, we address three key questions: first, is there good evidence for generalization in avoidance learning in humans?; second, can we distinguish behavioural and neural components relating to perceptual, aversive value, and safety value?; and third, which if any component predicts relevant psychological symptoms? We used a custom-designed perceptual task in conjunction with reinforcement learning modelling to study two groups: a laboratory-based sample (N = 26) who performed a pain avoidance task with concurrent neuroimaging (fMRI), and a larger cohort of individuals (N = 482), who performed a monetary loss avoidance task online alongside a battery of questionnaires designed to probe relevant psychological symptom dimensions (Gillan and Daw, 2016).

The overall study design is summarised in Figure 1a. In both groups of participants, generalization of instrumental responding was tested using a costly avoidance paradigm (Figure 1c). Briefly, participants were instructed that they would see a series of flower-like shapes on their screen, some of which were ‘safe’, and some of which were ‘dangerous’. If they saw a dangerous shape and made no response, there was a high chance that they would receive a painful electric shock (fMRI sample), or lose 10 cents from their cash stake (online sample using Amazon Mechanical Turk, AMT). If they saw a safe shape, they would never receive a shock (or lose money) on that trial. In order to escape the possibility of a painful shock (or monetary loss) when they thought a dangerous shape had been presented, participants were told they could press the ‘escape’ button on their keypad. Participants were instructed that the aversive outcome would never occur on a trial when they had pressed the ‘escape’ button – but – that, importantly, pressing the button was associated with a small cost. Specifically, each time they pressed the escape button, it would be registered on a counter at the bottom of their screen. At the end of each block of the task, they would receive additional painful shocks (or lose additional cash) depending on how many times they had pressed the button during that block (one extra shock or 10 cent loss per every five button presses). The optimal strategy (in order to minimise the amount of pain received or money lost) would therefore be to press the button if they thought they saw a dangerous shape, but not press if they thought a safe shape was on the screen.

Figure 1 with 3 supplements see all

Download asset Open asset

Crucially, on a small proportion of trials, the presented shapes were generalization stimuli (GSs). GSs were individually generated using precise estimates of perceptual ability (as measured on the first study session for the fMRI group) to be 75% reliably perceptually distinguishable from the task stimuli associated with aversive outcomes (CS+ s). (Due to time constraints and lack of control over testing environment, GS were generated based on average perceptual acuity from a pilot study in the online group.) The perceptual task (Figure 1b) was custom designed based on the recommendations of a recent review (Struyf et al., 2015). Specifically, in order to provide a fair test of perceptual performance during the generalization task, stimuli were not instantly comparable (in order to ensure that GSs would be reliably discriminable in an absolute sense, when presented in isolation; [Slivinske and Hall, 1960]), and testing occurred in the same emotional context (i.e. under threat of painful shock).

Importantly, the task stimulus array (in terms of arrangement of CS+ and CS- stimuli in perceptual space) was specifically chosen to probe asymmetries in generalization behaviour that result from value-based mechanisms – see Figure 1b. One such potential asymmetry is a characteristic shift in peak responding from the CS+ to surrounding GSs, away from the direction of the CS- in perceptual space (known as ‘peak shift’), that has been proposed to result from the interaction of excitatory and inhibitory generalization gradients around CS+ and CS- stimuli following Pavlovian conditioning (Hanson, 1959). Crucially, the asymmetric array used here allowed us to compare responses to CS+ GSs both near and far in perceptual space from the CS- – enabling detection of gradient interaction effects such as peak shift in instrumental avoidance, and allowing the separation of oppositely signed generalization gradients around CS+ and CS- stimuli.

We conducted a series of analyses on data from our two cohorts in order to address our key questions. First, we used reinforcement learning modelling to investigate whether there was evidence of value-based generalization in avoidance behaviour. Next, we used univariate fMRI data analysis to identify brain regions that encoded modelled internal quantities specific to value-based generalization processes. We then took a multivariate approach to investigate how the distributed representation of generalization stimuli in these regions changed over the course of the task, and how this related to individual differences in generalization. Finally, we used data from our online questionnaire battery to determine whether specific elements of avoidance generalization were related self-reported psychological symptoms.

Evidence for generalization in avoidance behaviour

For both groups of participants, the frequency of avoidance in response to generalization stimuli was intermediate to that evoked by CS- and CS+ stimuli (all p<0.0001, paired-sample t tests; fMRI: GS vs CS- t₂₅ = 7.57, mean difference = 0.18 [95%CI 0.14–0.24], GS vs CS+ t₂₅ = −17.6, mean difference = −0.60 [95%CI −0.67 to −0.54]; AMT: GS vs CS- t₄₈₁ = 27.0, mean difference = 0.35 [95%CI 0.33 – 0.38], GS vs CS+ t₄₈₁ = −26.6, mean difference = −0.20 [95%CI −0.19 to −0.21]; Figure 1d,e). Despite never having been associated with the aversive outcome, participants also rated GSs significantly higher than CS- (but lower than CS+) stimuli on post-task pain/loss expectancy scales (all p<0.0001, paired-sample t tests; fMRI: GS vs CS- t₂₅ = 5.69, mean difference = 24.1 [95%CI 15–33], GS vs CS+ t₂₅ = −8.14, mean difference = -52 [95% CI −39 to −66]; AMT: GS vs CS- t₄₈₁ = 29.4, mean difference = 41.7 [95%CI 40.0–44.6], GS vs CS+ t₄₈₁ = −16.5, mean different = -18 [95% CI −16.0 to −20.3], on visual analogue scales ranging 0–100; [Figure 1d,e]).

There was also a significant positive relationship between relative GS avoidance and relative GS pain/loss expectancy rating post-task in both groups (fMRI, Spearman’s ρ = 0.655, p=0.00027; AMT, Spearman’s ρ = 0.432, p=2.2e-16; both measures within-participant z-transformed, for relationships between raw scores see [Figure 1—figure supplement 1]). This suggests that a higher frequency of avoidance responding (plus associated lack of extinction) translated into higher conscious negative expectancy beliefs for generalization stimuli. There was no relationship between proportionate avoidance on GS trials and perceptual acuity at session 1 (individual θ values) or absolute intensity of the painful electrical stimulation (current amplitude) in the fMRI sample (all p>0.2).

This raises the question as to whether the observed avoidance on the GS trials was over and above that which would be expected from perceptual uncertainty alone. Notably, mean proportionate avoidance on GS trials in the fMRI group was around 0.2 (or ~0.25 when scaled relative to individual mean CS+ avoidance) – which, given that GSs were generated to be 75% reliably distinguishable from CS+s, is what might have been predicted from a purely perceptual account of task performance. Mean reaction times for making avoidance responses were also significantly slower for GS compared to CS+ stimuli in both groups, suggesting greater uncertainty on these trials (p=0.006, p=2.07e-11, paired sample t tests; fMRI: t₂₅ = 3.00, mean difference = 167 ms [95% CI 51.2–282], AMT: t₄₈₁ = 6.87, mean difference = 38.8 ms [95% CI 27.7–49.9]; [Figure 1d,e]). To resolve this issue, we tested for the presence of additional value-based generalization processes in both datasets using a principled model comparison approach.

Simply, we fitted a series of reinforcement learning models to avoidance data from both samples (modified Q-learning algorithms, with trial-by-trial varying learning rates determined by the Pearce-Hall associability rule, [Sutton and Barto, 1998; Le Pelley, 2004] – see Materials and methods). Firstly, we fit a model with perceptual ‘generalization’ only (modelled as 25% chance of perceptual confusion between GSs and the adjacent CS+) – that is where all task stimuli were treated as independent states, with no transfer of value across states. Secondly, we fit a model with perceptual generalization plus an additional value-based generalization process. As there is some evidence that generalization functions are approximately Gaussian in shape, at least along a single perceptual dimension (Ghirlanda and Enquist, 2003), this was implemented as a Gaussian smoothing of stimulus value across perceptual space, with a single free parameter (σ) governing the width of this function. Thirdly, we fit a model with perceptual generalization plus two additional free parameters governing width of additional value-based generalization processes – one for aversive (shock/loss) and one for neutral (no shock/no loss) feedback (σ_A and σ_N, respectively). This model was informed by previous empirical observations that generalization functions vary in gradient or width for aversive, neutral, and rewarding feedback (Schechtman et al., 2010; Resnik and Paz, 2015; Laufer et al., 2016).

The above models were fit to avoidance data from both groups using a variational Bayes approach to model inversion, under a mixed-effects framework (whereby within-subject priors are iteratively refined and matched to the inferred parent population distribution; see Materials and methods). Random-effects Bayesian model comparison indicated that in both samples the model with two additional value-generalization mechanisms (separately governing width of generalization from aversive and neutral feedback) best accounted for the avoidance data, as indexed by exceedance probability (probability that the model in question was the most frequently utilised in the population; fMRI, EP = 0.823, AMT, EP =~ 1; Figure 2a).

Figure 2

Download asset Open asset

For both fMRI and AMT data, this model provided a good account of avoidance decisions. Mean predictive accuracy (r², for binary choice data this is equivalent to the percentage of correct classifications) was 0.868 (±0.07) for fMRI and 0.849 (±0.11) for AMT groups, and the Bayesian ‘p value’ (posterior probability of the null hypothesis of random choice) was ≤6.8e-7 for all fMRI participants, and ≤0.026 for 477/482 AMT participants. In both groups, values of the parameter describing the width of aversive feedback (σ_A) were unrelated to values of other model parameters governing learning rate, choice bias, and choice stochasticity (see Materials and methods; all p>0.09), suggesting sufficient parameter identifiability. In both samples, σ_A values were significantly larger than values of the parameter governing width of generalization from neutral (safe) feedback, σ_N, indicating wider generalization for aversive compared to neutral outcomes (p=3.0e-8, p=2.2e-16, related-samples Wilcoxon signed rank tests; fMRI: mean σ_A=0.752 ± 0.29, mean σ_N=0.028 ± 0.03; AMT: mean σ_A=0.695 ± 0.23, mean σ_N=0.057 ± 0.05). Interestingly, σ_A values were not significantly related to σ_N values (fMRI group, Spearman’s ρ = −0.169, p>0.4; AMT group, ρ = 0.06, p>0.17), suggesting these may be at least partially independent processes.

Importantly, only a model including additional value-based generalization mechanisms can generate asymmetries in avoidance behaviour across pairs of generalization stimuli (peak shift), as apparent in Figure 1—figure supplement 2. Further, example traces for two representative participants from the fMRI group (Figure 2b) illustrate that stimulus values tend to asymptote – i.e. that under this model generalization of value across stimuli is assumed to be relatively constant over time. This assumption is consistent with our behavioural data, in that a time-on-task analysis showed that after initial period of exploratory learning (blocks 1–2), generalization in terms of GS avoidance remains fairly stable. In both groups of participants, there were significant effects of both CS type and block number, and a CS type*block interaction, on proportionate avoidance responding (fMRI: F_2,50=406.3, F_4,100=6.14, F_8,200=8.68, respectively; AMT: F_2,962=1077.9, F_4,1962=24.3, F_8,3848=263.0, respectively; all p<0.001, repeated-measures ANOVA). In the fMRI sample, the CS type*block interaction was driven by lower avoidance for CS+ stimuli in block one compared to the rest of the task (p≤0.004; other CS types no significant differences between blocks; pairwise comparisons Bonferroni corrected for multiple comparisons). This suggests a strategy of exploratory non-avoidance to enable proper learning of CS+ stimuli in block 1, but fairly constant generalization of avoidance across later blocks. In the AMT sample, there was also lower avoidance for CS+ stimuli in block one vs other blocks (all p<0.001), but a decrease in avoidance for CS- stimuli in later blocks (3-5) vs earlier blocks (1 and 2; all p<0.001). Overall GS avoidance showed small increases then decreases over first three blocks (p<0.001), before stabilising between blocks 4 and 5 (p>0.5, Bonferroni-corrected pairwise comparisons; see [Figure 1—figure supplement 2]).

Brain regions encoding model quantities specific to value-based generalization

As our behavioural data provided evidence for the presence of generalization in instrumental avoidance in both groups, we next employed a univariate analysis approach to our functional imaging data in order to investigate whether model quantities specific to value-related generalization processes were encoded in regional blood oxygen level-dependent (BOLD) signals.

In addition to work highlighting the role of the insula, amygdala, and primary sensory cortex in aversive generalization following Pavlovian conditioning (Ghosh and Chattarji, 2015; Onat and Büchel, 2015; Resnik and Paz, 2015; Laufer et al., 2016), previous functional imaging studies have identified the striatum and prefrontal cortex as encoding generalization gradients in healthy human volunteers (Dunsmoor et al., 2011; Greenberg et al., 2013; Lissek et al., 2014). However, the contribution of perceptual uncertainty (i.e. absolute discriminability of ‘generalization stimuli’ compared with other conditioned stimuli) is not always adequately addressed in the study of such gradients. Here, we used a strict parametric approach to identify additional variance in regional BOLD that can be attributed to our winning value-based generalization model, over and above that which can be explained by a purely perceptual account. This was achieved by using serially orthogonalised regressors derived from each model to predict trial-by-trial variation in BOLD signal in our regions of interest (see Figure 3a and Materials and methods).

Figure 3

Download asset Open asset

We found evidence for the encoding of additional variance in trial-by-trial expected stimulus values derived from the value-based generalization model in both the anterior insular cortex and the dorsal striatum (Figure 3b). BOLD signal was greater when the expected value of a particular stimulus was lower (or the predicted probability of receiving a painful shock if an avoidance response was not made was higher) in the left anterior insula (p_WB = 0.0073, k = 73, peak voxel [−30,23,–4], Z = 4.71; sub-threshold trend in the right anterior insula: p_SVC = 0.073, k = 9, peak voxel [42,23,-1], Z = 3.45), and right caudate (p_SVC = 0.024, k = 20, peak voxel [9,8,8], Z = 3.95). There was no evidence for univariate encoding of this signal in primary visual cortex (V1) or the amygdala. We also found no evidence for negative encoding of aversive value (greater BOLD signal with lower predicted probability of shock, or ‘safety signalling’) in the ventromedial prefrontal cortex (vmPFC).

In addition to expected value signals, we examined potential encoding of prediction errors, which are the main learning signals in reinforcement learning (PEs; defined as the difference between actual and predicted outcome on any given trial – see Materials and methods). We focused our analysis on negatively signed PEs (generated on trials where no shock was received, but the predicted P(shock) was >0), as this both constrains analysis to trials where an avoidance response was not made (on avoidance trials PE = 0, by definition), and gives greater weighting to generalization trials where, due to perceptual uncertainty alone, predicted P(shock) will be >0, but no aversive outcome is ever delivered. (Positively signed PEs are highly collinear with shock administration and therefore are hard to detect under our design.)

We also found evidence of significant encoding of additional variance in PE signals from the value-based generalization model in insula and striatum (Figure 3c). Specifically, BOLD signal was greater when trial PE was more negative in the anterior insula, bilaterally (left: p_SVC = 9.72e-5, k = 93, peak voxel [−33,20,11], Z = 5.48; right: p_SVC = 0.024, k = 19, peak voxel [33,26,-4], Z = 4.35), right insula more posteriorly (p_SVC = 5.85e-5, k = 65, peak voxel [48,8,-4], Z = 4.40), putamen, bilaterally (left: p_SVC = 0.024, k = 20, peak voxel [−27,–4,−1], Z = 4.29; right: p_SVC = 0.009, k = 31, peak voxel [33,2,-1], Z = 4.06), and right pallidum (p_SVC = 0.046, k = 14, peak voxel [18,5,2], Z = 3.74). Significant clusters were also observed in the mid cingulate cortex (p_WB = 0.001, k = 103, peak voxel [6,14,44], Z = 4.46), left parietal operculum (p_WB = 3.56e-5, k = 168, peak voxel [−48,–25,14], Z = 4.10), right inferior parietal lobule (p_WB = 0.003, k = 90, peak voxel [54,-40,26], Z = 3.82) and inferior frontal gyrus (p_WB = 0.023, k = 56, peak voxel [42,5,35], Z = 4.31) − but we found no evidence of encoding of value generalization-derived PE signals in V1, the amygdala, or vmPFC.

Changes in neural representation of generalization stimuli over the course of the task: relationship to individual differences in avoidance behaviour

Previous studies in animal models have shown that over the course of conditioning, the representation of the conditioned stimulus (CS+) in terms of response pattern across many individual units may come to resemble that of the primary aversive reinforcer (e.g. Grewe et al., 2017). To complement our univariate results, we therefore examined how different task stimuli were represented in multivariate space using representational similarity analysis (Kriegeskorte et al., 2008). This approach enables the consideration of the full representational geometry across specific brain regions – how information is encoded, as well as whether or not it is – and depends on the calculation of distance metrics to quantify how (dis)similarly different kinds of stimuli are represented in multivariate space (in fMRI, across all voxels in a particular brain volume).

Following the approach of a recent study of aversive conditioning in rodents (Grewe et al., 2017), we examined how representational difference changed in our regions of interest earlier (blocks 1–2) vs later (blocks 3–5) in the task – and, crucially, how this change related to individual differences in overall behavioural expressions of conditioning. Specifically, we investigated whether changes in representation of GS, relative to CS+, stimuli over the course of the task related to individual tendency to generalize value from CS+ to GS stimuli – as captured behaviourally in avoidance responses on GS trials. We calculated a robust, cross-validated estimate of representational distance, Fisher’s linear discriminant contrast (see Materials and methods, Figure 4a) in order to maximise the reliability of our results. Importantly, the use of a cross-validated distance measure means that derived (dis)-similarity estimates are unbiased by noise (which may potentially vary across individuals and imaging runs), and have a meaningful zero point (Walther et al., 2016).

Figure 4 with 1 supplement see all

Download asset Open asset

Overall, for no region of interest was there a significant group level change in representational distance between GS and CS+ stimuli (all p>0.03, paired-sample t tests; Bonferroni-corrected threshold = 0.01 for alpha = 0.05). However, across individuals, greater increase in similarity of representation of GS to CS+ stimuli over the course of the task in primary visual cortex was related to greater behavioural generalization in terms of greater relative GS avoidance (p=0.010, multiple linear regression model; Table 1, Figure 4b). For individuals who made a higher relative proportion of avoidance responses towards generalization stimuli, V1 representation of GS stimuli came to be more similar to that of CS+ stimuli over the course of the task – but for individuals who avoided less on GS trials, GS stimuli came to be less similarly represented to CS+s in these regions (for visualisation of the relationship between raw proportionate GS avoidance and V1 distance change, see Figure 4d). There was no evidence of a significant relationship between GS−CS+ representational distance change and relative GS avoidance in the anterior insula, striatum, amygdala or vmPFC (Table 1, Figure 4b). We confirmed these results by implementing a cross-validated regularised regression (CV LASSO, see Materials and methods) on the same data (this kind of regression shrinks non-significant predictor coefficients to zero, and generally results in smaller coefficients compared to traditional linear regression). Under this robust approach, change in GS−CS+ similarity in V1, but not other regions, was retained as a significant predictor of relative GS avoidance (β = −0.040), in the model that minimised mean squared error (MSE).

Table 1

Change in GS–CS+ representational distance	β	SE	t	p
a. insula	−0.04287	0.06798	−0.631	0.535
caudate	−0.02304	0.04173	−0.552	0.587
amygdala	−0.09792	0.09905	−0.989	0.335
V1	−0.10072	0.03531	−2.852	0.010*
vmPFC	−0.07407	0.07938	−0.933	0.362

Using a post hoc test, we examined whether changes in GS−CS+ representational distance in V1 might relate to changes in absolute discriminability of generalization stimuli (as measured on the day before and day after the generalization test session). Mean discriminability for GSs (CS+ ± θ) was 0.75 on session 1, by definition, and 0.79 on session 3 (±0.14, range 0.465–0.994; although note at the group level there was no significant change in θ values measured across sessions, see above). Under this exploratory analysis, we found evidence of a significant association between change in V1 GS−CS+ representational distance during the task, and post-conditioning changes in perceptual discriminability of the GSs. Individuals who showed an increase in similarity of representation showed worse perceptual performance post-(vs pre-) conditioning, and those who showed decreased similarity showing better performance (Spearman’s ρ = 0.518, p=0.007; see [Figure 4—figure supplement 1]). There was no significant relationship between change in perceptual acuity and representational distance in any other brain region (all p>0.09).

All the univariate fMRI findings presented above remained significant if re-ran using regressors derived from a model where perceptual discriminability of GSs changes linearly over the course of the task from pre- to post-conditioning measured acuity levels (full, unthresholded statistical maps for all analyses are available at Neurovault; neurovault.org/collections/3177).

Relationship between individual differences in value-based generalization and self-reported psychopathology

Hypotheses about the role of generalization in psychological disorders tend to relate to an over-generalization of aversive information – but it has also been proposed that poor discrimination (e.g. between CS+ and CS- in anxiety groups) may be due to inadequate learning about safety cues. We therefore looked first at how psychological symptoms scores related to individual σ_A values, but also examined possible relationships with individual σ_N values, in our online cohort (N = 482).

Following the approach of Gillan et al. (2016), the online group completed a battery of self-report questionnaires that probed symptoms hypothesized to be related to aversive over-generalization (trait anxiety, mood disorder symptoms, obsessive-compulsive traits, and ‘global’ cognitive style), in addition to some positive control measures (apathy and impulsivity scales). (A summary of scores on these measures and other demographic information for both samples is available in Supplementary file 1). To enable comparison with the findings of Gillan et al., self-report information was first compared to individual parameter estimates using precision-weighted linear regression models, controlling for age and gender identity (see Materials and methods). This approach was then complemented by the implementation of cross-validated regularised regression models (CV LASSO regression), as in the previous section (these models also included age and gender identity as regressors of no interest).

First, we sought to identify whether individual values of the parameter governing width of generalization from aversive feedback (σ_A) were related to symptom scores on any measure. Total scores across measures exhibited good to excellent internal reliability (mean Cronbach’s α = 0.882, see Supplementary file 2), and, as might be expected, covaried significantly across participants (mean absolute r for inter-correlation between scores = 0.479). Regression of total scores against parameter estimates was therefore implemented in separate models for each measure, in order to enable meaningful partition of variance. The Nyholt-Bonferroni corrected p value for significance across these separate models of non-independent measures was p<0.010 to maintain an alpha of 0.05 (effective number of independent variables = 5.0, see Materials and methods).

Parameter estimates governing width of generalization from aversive feedback were found to be significantly positively associated with trait anxiety scores (greater width with greater anxiety), and significantly negatively associated with trait apathy (smaller width with greater apathy; anxiety, p=0.009, apathy, p<0.001, individual precision-weighted linear regression models controlling for age and gender; see Table 3, Figure 5a). These two effects remained significant when trait anxiety and apathy scores were included in the same model, suggesting they were independent (anxiety: β = 0.050, SE 0.015, t = 3.34, apathy: β = −0.060, SE 0.014, t = −4.28; both p<0.001). This result was confirmed under the cross-validated and regularised analysis; when all predictors were entered in the same model both anxiety and apathy total scores were retained as predictors in the model that minimised MSE (β = 0.021, β = −0.032, respectively). No questionnaire total scores were significantly related to σ_N values (p>0.05).

Table 3

Questionnaire measure	β	SE	t	p
STAI total	0.039	0.015	2.626	0.009*
AMI total	−0.051	0.014	−3.687	<0.001*
OCI-R total	0.005	0.014	0.373	0.710
PHQ9 total	0.021	0.015	1.476	0.141
BIS-11 total	−0.005	0.013	−0.410	0.682
CSQ global	−0.014	0.014	−0.978	0.328

Figure 5

Download asset Open asset

As per Gillan et al, we also sought to reduce collinearity in our battery of self-report measures by entering all recorded items (N = 142) into a factor analysis. Using an identical method to that described in the previously cited paper (see Materials and methods), we derived a three-factor solution (for scree plot see Figure 5b). These factors were labelled ‘intrusive anxiety’, ‘low self-worth’, and ‘low self-control’ on the basis of their top loading items (see Figure 5c).

The ‘intrusive anxiety’ factor was mostly composed of items from the trait scale of State-Trait Anxiety Inventory (STAI; 20 items, mean loading = 0.457 ± 0.12), Obsessive-Compulsive Index (OCI; 18 items, mainly items probing intrusive thoughts and checking behaviour, mean loading = 0.602 ± 0.087), Physician’s Health Questionnaire (PHQ9; eight items probing mood disorder symptoms, mean loading = 0.531 ± 0.056), and the Barratt Impulsivity Scale (BIS; six items pertaining to racing/intrusive thoughts and restlessness, mean loading = 0.386 ± 0.15). ‘Low self-worth’ was mostly comprised of items from the Cognitive Style Questionnaire (CSQ; 37 items, mainly from low self-worth and internal attribution subscales, mean loading = 0.518 ± 0.13) and the STAI (11 items, mainly related to low self-worth/negative self-affect, mean loading = 0.322 ± 0.054). 'Low self-control' mostly comprised items from the BIS (23 items, mainly from the non-planning and attentional impulsivity subscales, mean loading = 0.485 ± 0.15), with some loading from the apathy motivation index (AMI; six items from the behavioural amotivation subscale, mean loading = 0.356 ± 0.093) and STAI (seven items relating to feel uncontent/unrested, mean loading 0.321 ± 0.04). (For full item loadings for each factor, see Supplementary file 3).

The ‘intrusive anxiety’ factor analysis-derived symptom score was significantly and selectively related to individual differences in aversive generalisation width (σ_A values) – in both multiple linear and robust regression models (p=0.008, precision-weighted multiple regression model; see Table 4, Figure 5c; only factor retained in MSE-minimising CV LASSO model, β = 0.019). None of the factor analysis-derived symptom scores were related to individual σ_N values (all p>0.1).

Table 4

Factor analysis-derived symptom score	β	SE	t	p
‘Intrusive anxiety’	0.043	0.016	2.677	0.008*
‘Low self-worth’	−0.019	0.015	−1.255	0.210
‘Lack of self-control’	−0.000	0.014	−0.032	0.975

The results presented here provide robust evidence for generalization in human avoidance learning. In particular, we demonstrate that generalization involves a number of distinct processes relating to different components of avoidance: perceptual uncertainty, aversive value generalization, and neutral (safety) value generalization. These processes each relate to different patterns of neural representations in the brain. Finally, we show that aversive value generalization is a specific predictor of trait anxiety in a large population sample.

Examining instrumental avoidance behaviour allows us to investigate how individuals learn about and attribute value to the set of actions they can take when faced with a particular stimulus or situation (as distinct from passively learnt Pavlovian stimulus-value associations). Using reinforcement learning modelling, we found behavioural evidence for additional value-based contributions to avoidance generalization (i.e. over and above that which might be expected from perceptual uncertainty alone) in two independent groups of participants (sampling different populations, and using two different kinds of aversive reinforcer). Notably, choice data from both groups supported an account of value-generalization that allowed for different widths of generalization from aversive (pain or monetary loss) vs neutral (no pain or loss) feedback. Consistent with previous evidence from studies of generalization of Pavlovian conditioning in humans and non-human primates, we observed larger width generalization functions for aversive compared to neutral feedback (Schechtman et al., 2010; Resnik and Paz, 2015; Laufer et al., 2016). In both groups, estimates of free parameters governing widths of these two processes were uncorrelated, suggesting they might relate to at least partially separable mechanisms.

Taking an explicit model-based approach enabled us to identify brain regions where BOLD signal was related to variance in modelled quantities specific to value-based generalization (namely, expected value and prediction error signals). When potential perceptual confusion between visually similar task stimuli was properly accounted for, we found evidence for encoding of value-related generalization signals in the anterior insula and dorsal striatum. The anterior insula and striatum (more ventrally) have previously been implicated in representing expected value and prediction error signals in higher-order pain conditioning (Seymour et al., 2004), and the dorsal striatum is implicated in prediction error signals in avoidance learning (Palminteri et al., 2012; Seymour et al., 2012; Eldar et al., 2016), suggesting an important role for these structures in aversive learning (see also Delgado et al., 2009). Dorsal, rather than more ventral striatal control has also been implicated in the transfer from goal-directed to habit-based avoidance in instrumental paradigms (LeDoux et al., 2017). Greater understanding of habitual control in excessive avoidance has particular clinical relevance as it may explain why maladaptive avoidance can persist following extinction (e.g. contributing to treatment-resistance in exposure therapy for anxiety disorders, [Treanor and Barry, 2017]), and has been proposed as core mechanism in obsessive-compulsive disorder (Gillan et al., 2014). We found no evidence of univariate encoding specific to value-based model quantities in the amygdala, primary visual cortex (V1), or ventromedial prefrontal cortex (vmPFC). However, this may be because this kind of analysis is not ideally suited to detect distributed representations involved in associative learning.

In previous studies of Pavlovian aversive conditioning, it has been demonstrated that positively conditioned stimuli come to be more closely represented to the primary aversive outcome in multivariate space (e.g. across neural ensemble activity in the basolateral amygdala, [Grewe et al., 2017]). Here, we used a robust, cross-validated measure of representational distance to analyse data across all voxels in regions of interest, and found that increased similarity of representation of GS to CS +stimuli over the course of the task in primary sensory cortex was related to higher overall behavioural generalization (higher proportionate avoidance on generalization trials). Individuals for whom GS stimuli came to be more closely represented to CS +s in these brain regions (despite never having been directly associated with the aversive outcome) chose to avoid more in the face of GS stimuli − and vice versa. This change in representational geometry, in association with the lack of opportunity for extinction of inappropriately generalized value in an avoidance context, may have contributed towards the stability of generalization (in terms of overall GS avoidance) we observed over the later phases of the task.

Consistent with perceptual accounts of generalization, a post-hoc analysis suggested that representational change for GSs relative to CS +stimuli over the course of the task in primary visual cortex might account for some of the generalisation in avoidance we observed (in addition or parallel to value-based mechanisms identified above). Individuals who avoided more frequently on generalization trials, and who showed associated increases in GS−CS +representational similarity in V1, exhibited decreased perceptual acuity for task stimuli on next day perceptual testing - with the opposite pattern observed in participants who showed lower GS avoidance. Absolute decreases in discriminability for task stimuli result in increased generalization ‘for free’ (without having to involve additional mechanisms), and therefore may contribute to maintenance of generalization in some participants.

However, consistent with accounts that favour the involvement of a wider network of brain regions in coordinating generalization across stimuli, we also found a role for multivariate anterior insula and amygdalar representations in individual differences in aversive value generalization. Individuals who had higher estimates for the model parameter governing value generalization specifically from aversive feedback showed greater increases in similarity of GS−CS +representation in these regions. Somewhat surprisingly, the opposite relationship was observed in primary visual cortex, such that increases GS-CS +similarity in this region were associated with lower individual aversive generalization parameter estimates. One potential explanation for this finding is that some kind of compensatory mechanism exists between perceptual and value-based generalization processes, acting at the brain rather than behaviour level. Interestingly, changes in discriminative ability following aversive conditioning have recently been associated with altered insula and amygdalar processing of visual stimuli in humans (Shalev et al., 2018). However, this result was unexpected and would therefore benefit considerably from further investigation in future work.

Although less well optimised under our design, we also conducted an analysis to probe whether changes in GS relative to CS- stimuli might be associated with individual estimates of the model parameter governing width of generalization specifically from neutral (or ‘safe’) outcomes (in this case, omission of painful shock). Individuals with higher values of the parameter governing extent of generalization from neutral feedback exhibited greater increases in GS−CS- similarity over the course of the task in both the amygdala and vmPFC. This adds to a body of work suggesting that amygdalar function is not only important for the generalization of fear responses, but that it is also involved in safety learning (Genud-Gabai et al., 2013; Likhtik et al., 2014). A recent study in rodents suggests that the lateral amygdala may be particularly important region for understanding individual differences in fear behaviour towards perceptually ambiguous novel stimuli, with different neuronal sub-populations involved in successful discrimination of novel safe stimuli and inappropriate fear responses – in a way that would be hard to detect by averaging signal across this region as a whole (Grosso et al., 2018). Although the vmPFC has previously been demonstrated to show inverse perceptual similarity-derived generalization gradients following aversive conditioning (e.g. Lissek et al., 2014; Onat and Büchel, 2015), it is not always clear from the experimental design whether this represents the simple inverse of aversive gradients (stemming from the CS+), or rather the positive signalling of safety gradients (stemming from the CS-). The evidence presented here provides tentative support for the latter account, at least in an instrumental context.

Excessive avoidance in response to contexts or stimuli which do not pose a threat to an individual’s health or well-being can significantly impair general functioning and is often associated with high levels of psychological distress (Arnaudova et al., 2017). Such maladaptive avoidance has been identified as a core pathological dimension across several psychological disorders, including anxiety disorders, obsessive-compulsive disorder, chronic pain, and depression (LeDoux et al., 2017). Over-generalization of aversive feedback to encompass non-threatening but psychologically similar stimuli or contexts has been proposed as a key mechanism underlying the initiation and maintenance of excessive avoidance in these conditions (Duits et al., 2015; Dymond et al., 2015; Harvie et al., 2017; Pearson et al., 2015) – however, the link between generalization of negative value and inappropriate avoidance behaviour has been relatively underexplored.

We found selective relationships between psychological symptom scores and individual parameter estimates governing width of value generalization from aversive, but not safe/neutral outcomes. The largest positive relationship between symptom score and magnitude of aversive generalisation was for the factor-analysis derived score labelled ‘intrusive anxiety’, which mainly comprised items probing self-reported trait anxiety, but also reports of intrusive thoughts from the obsessive-compulsive inventory (% increase in parameter value with a 1SD increase in symptom score was 11.0% for intrusive anxiety, and 10.6% for trait anxiety alone). We also found a significant negative relationship between self-reported apathy and aversive generalization (22.9% decrease in parameter value with a 1SD increase in symptom score) – an effect which appeared to be independent from that relating to self-reported anxiety. This is an interesting finding, as we often think about apathy as involving a greater sensitivity to perceived effort, or decreased sensitivity to potential rewards, rather than a decreased impact of information about punishments (e.g. Bonnelle et al., 2015). As, to our knowledge, there has been no previously reported link between self-reported apathy and aversive generalization, this finding would benefit from future replication.

In summary, the findings reported here demonstrate the benefits of parsing complex processes such as generalization into separate components, and examining individual relationships between these components and both neural mechanisms and self-reported psychopathology. This approach may help unify previous apparently contradictory observations, and underlines that both perceptual and value-based processes are likely at work in generalization phenomena. Identification of patients across diagnostic categories who may have a primary deficit in excessive aversive generalization may help target them towards treatments which work more effectively. Further, greater understanding of the mechanisms of over-generalization of avoidance (including transfer to habit-based control systems) may help improve understanding of treatment resistance in these disorders.

All relevant code for stimulus generation, data collection, and data analysis, in addition to raw behavioural data, is available at the project's Open Science Framework page (osf.io/25t3f). Raw functional imaging data is deposited at openfMRI (openfmri.org/dataset/ds000249) and derived statistical maps are available at NeuroVault (neurovault.org/collections/3177).

1. 1. Agnes Norbury

   (2017) Value generalization in human avoidance learning

   Publicly available at Neurovault (accession no. 3177).

   https://neurovault.org/collections/3177/
2. 1. Agnes Norbury

   (2017) Value generalization in human avoidance learning

   Publicly available at the Open Science Framework.

   https://osf.io/25t3f/
3. 1. Agnes Norbury
   2. Ben Seymour

   (2018) Value generalization in human avoidance learning

   Publicly available at Open Neuro.

   https://openneuro.org/datasets/ds000249/versions/00002

1. 1. Aizenberg M
   2. Geffen MN

   (2013) Bidirectional effects of aversive learning on perceptual acuity are mediated by the sensory cortex

   Nature Neuroscience 16:994–996.

   https://doi.org/10.1038/nn.3443

   • PubMed
   • Google Scholar
2. 1. Ang YS
   2. Lockwood P
   3. Apps MA
   4. Muhammed K
   5. Husain M

   (2017) Distinct subtypes of apathy revealed by the apathy motivation index

   Plos One 12:e0169938.

   https://doi.org/10.1371/journal.pone.0169938

   • PubMed
   • Google Scholar
3. 1. Arnaudova I
   2. Kindt M
   3. Fanselow M
   4. Beckers T

   (2017) Pathways towards the proliferation of avoidance in anxiety and implications for treatment

   Behaviour Research and Therapy 96:3–13.

   https://doi.org/10.1016/j.brat.2017.04.004

   • PubMed
   • Google Scholar
4. 1. Bonnelle V
   2. Veromann KR
   3. Burnett Heyes S
   4. Lo Sterzo E
   5. Manohar S
   6. Husain M

   (2015) Characterization of reward and effort mechanisms in apathy

   Journal of Physiology-Paris 109:16–26.

   https://doi.org/10.1016/j.jphysparis.2014.04.002

   • PubMed
   • Google Scholar
5. Book

   1. Carroll RJ
   2. Ruppert D

   (1988) Transformation and Weighting in Regression

   CRC Press.

   https://doi.org/10.1007/978-1-4899-2873-3

   • Google Scholar
6. 1. Cha J
   2. Carlson JM
   3. Dedora DJ
   4. Greenberg T
   5. Proudfit GH
   6. Mujica-Parodi LR

   (2014) Hyper-reactive human ventral tegmental area and aberrant mesocorticolimbic connectivity in overgeneralization of fear in generalized anxiety disorder

   Journal of Neuroscience 34:5855–5860.

   https://doi.org/10.1523/JNEUROSCI.4868-13.2014

   • PubMed
   • Google Scholar
7. 1. Crump MJ
   2. McDonnell JV
   3. Gureckis TM

   (2013) Evaluating Amazon's mechanical turk as a tool for experimental behavioral research

   PLoS ONE 8:e57410.

   https://doi.org/10.1371/journal.pone.0057410

   • PubMed
   • Google Scholar
8. 1. Daunizeau J
   2. Adam V
   3. Rigoux L

   (2014) VBA: a probabilistic treatment of nonlinear models for neurobiological and behavioural data

   PLoS Computational Biology 10:e1003441.

   https://doi.org/10.1371/journal.pcbi.1003441

   • PubMed
   • Google Scholar
9. 1. de Leeuw JR

   (2015) jsPsych: a JavaScript library for creating behavioral experiments in a Web browser

   Behavior Research Methods 47:1–12.

   https://doi.org/10.3758/s13428-014-0458-y

   • PubMed
   • Google Scholar
10. 1. Delgado MR
    2. Jou RL
    3. Ledoux JE
    4. Phelps EA

    (2009) Avoiding negative outcomes: tracking the mechanisms of avoidance learning in humans during fear conditioning

    Frontiers in Behavioral Neuroscience 3:33.

    https://doi.org/10.3389/neuro.08.033.2009

    • PubMed
    • Google Scholar
11. 1. Duits P
    2. Cath DC
    3. Lissek S
    4. Hox JJ
    5. Hamm AO
    6. Engelhard IM
    7. van den Hout MA
    8. Baas JM

    (2015) Updated meta-analysis of classical fear conditioning in the anxiety disorders

    Depression and Anxiety 32:239–253.

    https://doi.org/10.1002/da.22353

    • PubMed
    • Google Scholar
12. 1. Dunsmoor JE
    2. Prince SE
    3. Murty VP
    4. Kragel PA
    5. LaBar KS

    (2011) Neurobehavioral mechanisms of human fear generalization

    NeuroImage 55:1878–1888.

    https://doi.org/10.1016/j.neuroimage.2011.01.041

    • PubMed
    • Google Scholar
13. 1. Dymond S
    2. Dunsmoor JE
    3. Vervliet B
    4. Roche B
    5. Hermans D

    (2015) Fear generalization in humans: systematic review and implications for anxiety disorder research

    Behavior Therapy 46:561–582.

    https://doi.org/10.1016/j.beth.2014.10.001

    • PubMed
    • Google Scholar
14. 1. Eldar E
    2. Hauser TU
    3. Dayan P
    4. Dolan RJ

    (2016) Striatal structure and function predict individual biases in learning to avoid pain

    PNAS 113:4812–4817.

    https://doi.org/10.1073/pnas.1519829113

    • PubMed
    • Google Scholar
15. 1. Foa EB
    2. Huppert JD
    3. Leiberg S
    4. Langner R
    5. Kichic R
    6. Hajcak G
    7. Salkovskis PM

    (2002) The obsessive-compulsive inventory: development and validation of a short version

    Psychological Assessment 14:485–496.

    https://doi.org/10.1037/1040-3590.14.4.485

    • PubMed
    • Google Scholar
16. 1. Fullana MA
    2. Harrison BJ
    3. Soriano-Mas C
    4. Vervliet B
    5. Cardoner N
    6. Àvila-Parcet A
    7. Radua J

    (2016) Neural signatures of human fear conditioning: an updated and extended meta-analysis of fMRI studies

    Molecular Psychiatry 21:500–508.

    https://doi.org/10.1038/mp.2015.88

    • PubMed
    • Google Scholar
17. 1. Genud-Gabai R
    2. Klavir O
    3. Paz R

    (2013) Safety signals in the primate amygdala

    Journal of Neuroscience 33:17986–17994.

    https://doi.org/10.1523/JNEUROSCI.1539-13.2013

    • PubMed
    • Google Scholar
18. 1. Ghirlanda S
    2. Enquist M

    (2003) A century of generalization

    Animal Behaviour 66:15–36.

    https://doi.org/10.1006/anbe.2003.2174

    • Google Scholar
19. 1. Ghosh S
    2. Chattarji S

    (2015) Neuronal encoding of the switch from specific to generalized fear

    Nature Neuroscience 18:112–120.

    https://doi.org/10.1038/nn.3888

    • PubMed
    • Google Scholar
20. 1. Gillan CM
    2. Daw ND

    (2016) Taking Psychiatry Research Online

    Neuron 91:19–23.

    https://doi.org/10.1016/j.neuron.2016.06.002

    • PubMed
    • Google Scholar
21. 1. Gillan CM
    2. Kosinski M
    3. Whelan R
    4. Phelps EA
    5. Daw ND

    (2016) Characterizing a psychiatric symptom dimension related to deficits in goal-directed control

    eLife 5:e11305.

    https://doi.org/10.7554/eLife.11305

    • PubMed
    • Google Scholar
22. 1. Gillan CM
    2. Morein-Zamir S
    3. Urcelay GP
    4. Sule A
    5. Voon V
    6. Apergis-Schoute AM
    7. Fineberg NA
    8. Sahakian BJ
    9. Robbins TW

    (2014) Enhanced avoidance habits in obsessive-compulsive disorder

    Biological Psychiatry 75:631–638.

    https://doi.org/10.1016/j.biopsych.2013.02.002

    • PubMed
    • Google Scholar
23. 1. Glover GH
    2. Li TQ
    3. Ress D

    (2000) Image-based method for retrospective correction of physiological motion effects in fMRI: RETROICOR

    Magnetic Resonance in Medicine 44:162–167.

    https://doi.org/10.1002/1522-2594(200007)44:1<162::AID-MRM23>3.0.CO;2-E

    • PubMed
    • Google Scholar
24. 1. Gorgolewski KJ
    2. Auer T
    3. Calhoun VD
    4. Craddock RC
    5. Das S
    6. Duff EP
    7. Flandin G
    8. Ghosh SS
    9. Glatard T
    10. Halchenko YO
    11. Handwerker DA
    12. Hanke M
    13. Keator D
    14. Li X
    15. Michael Z
    16. Maumet C
    17. Nichols BN
    18. Nichols TE
    19. Pellman J
    20. Poline JB
    21. Rokem A
    22. Schaefer G
    23. Sochat V
    24. Triplett W
    25. Turner JA
    26. Varoquaux G
    27. Poldrack RA

    (2016) The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments

    Scientific Data 3:160044.

    https://doi.org/10.1038/sdata.2016.44

    • PubMed
    • Google Scholar
25. Conference

    1. Gorsuch RL
    2. Nelson J

    (1981)

    CNG scree test: an objective procedure for determining the number of factors

    Annual Meeting of the Society of Multivariate Experimental Psychology.

    • Google Scholar
26. 1. Greenberg T
    2. Carlson JM
    3. Cha J
    4. Hajcak G
    5. Mujica-Parodi LR

    (2013) Neural reactivity tracks fear generalization gradients

    Biological Psychology 92:2–8.

    https://doi.org/10.1016/j.biopsycho.2011.12.007

    • PubMed
    • Google Scholar
27. 1. Grewe BF
    2. Gründemann J
    3. Kitch LJ
    4. Lecoq JA
    5. Parker JG
    6. Marshall JD
    7. Larkin MC
    8. Jercog PE
    9. Grenier F
    10. Li JZ
    11. Lüthi A
    12. Schnitzer MJ

    (2017) Neural ensemble dynamics underlying a long-term associative memory

    Nature 543:670–675.

    https://doi.org/10.1038/nature21682

    • PubMed
    • Google Scholar
28. 1. Grosso A
    2. Santoni G
    3. Manassero E
    4. Renna A
    5. Sacchetti B

    (2018) A neuronal basis for fear discrimination in the lateral amygdala

    Nature Communications 9:1214.

    https://doi.org/10.1038/s41467-018-03682-2

    • PubMed
    • Google Scholar
29. 1. Grupe DW
    2. Nitschke JB

    (2013) Uncertainty and anticipation in anxiety: an integrated neurobiological and psychological perspective

    Nature Reviews Neuroscience 14:488–501.

    https://doi.org/10.1038/nrn3524

    • PubMed
    • Google Scholar
30. 1. Gureckis TM
    2. Martin J
    3. McDonnell J
    4. Rich AS
    5. Markant D
    6. Coenen A
    7. Halpern D
    8. Hamrick JB
    9. Chan P

    (2016) psiTurk: An open-source framework for conducting replicable behavioral experiments online

    Behavior Research Methods 48:829–842.

    https://doi.org/10.3758/s13428-015-0642-8

    • PubMed
    • Google Scholar
31. 1. Hanson HM

    (1959) Effects of discrimination training on stimulus generalization

    Journal of Experimental Psychology 58:321.

    https://doi.org/10.1037/h0042606

    • PubMed
    • Google Scholar
32. 1. Harvie DS
    2. Moseley GL
    3. Hillier SL
    4. Meulders A

    (2017) Classical conditioning differences associated with chronic pain: a systematic review

    The Journal of Pain 18:889–898.

    https://doi.org/10.1016/j.jpain.2017.02.430

    • PubMed
    • Google Scholar
33. 1. Kahnt T
    2. Park SQ
    3. Burke CJ
    4. Tobler PN

    (2012) How glitter relates to gold: similarity-dependent reward prediction errors in the human striatum

    Journal of Neuroscience 32:16521–16529.

    https://doi.org/10.1523/JNEUROSCI.2383-12.2012

    • PubMed
    • Google Scholar
34. 1. Kasper L
    2. Bollmann S
    3. Diaconescu AO
    4. Hutton C
    5. Heinzle J
    6. Iglesias S
    7. Hauser TU
    8. Sebold M
    9. Manjaly ZM
    10. Pruessmann KP
    11. Stephan KE

    (2017) The PhysIO toolbox for modeling physiological noise in fMRI Data

    Journal of Neuroscience Methods 276:56–72.

    https://doi.org/10.1016/j.jneumeth.2016.10.019

    • PubMed
    • Google Scholar
35. 1. Kriegeskorte N
    2. Formisano E
    3. Sorger B
    4. Goebel R

    (2007) Individual faces elicit distinct response patterns in human anterior temporal cortex

    PNAS 104:20600–20605.

    https://doi.org/10.1073/pnas.0705654104

    • PubMed
    • Google Scholar
36. 1. Kriegeskorte N
    2. Mur M
    3. Bandettini P

    (2008) Representational similarity analysis - connecting the branches of systems neuroscience

    Frontiers in Systems Neuroscience 2:4.

    https://doi.org/10.3389/neuro.06.004.2008

    • PubMed
    • Google Scholar
37. 1. Krypotos AM
    2. Effting M
    3. Kindt M
    4. Beckers T

    (2015) Avoidance learning: a review of theoretical models and recent developments

    Frontiers in Behavioral Neuroscience 9:189.

    https://doi.org/10.3389/fnbeh.2015.00189

    • PubMed
    • Google Scholar
38. 1. Laufer O
    2. Israeli D
    3. Paz R

    (2016) Behavioral and neural mechanisms of overgeneralization in anxiety

    Current Biology 26:713–722.

    https://doi.org/10.1016/j.cub.2016.01.023

    • PubMed
    • Google Scholar
39. 1. Laufer O
    2. Paz R

    (2012) Monetary loss alters perceptual thresholds and compromises future decisions via amygdala and prefrontal networks

    Journal of Neuroscience 32:6304–6311.

    https://doi.org/10.1523/JNEUROSCI.6281-11.2012

    • PubMed
    • Google Scholar
40. 1. Le Pelley ME

    (2004) The role of associative history in models of associative learning: a selective review and a hybrid model

    The Quarterly Journal of Experimental Psychology Section B 57:193–243.

    https://doi.org/10.1080/02724990344000141

    • Google Scholar
41. 1. LeDoux JE
    2. Moscarello J
    3. Sears R
    4. Campese V

    (2017) The birth, death and resurrection of avoidance: a reconceptualization of a troubled paradigm

    Molecular Psychiatry 22:24–36.

    https://doi.org/10.1038/mp.2016.166

    • PubMed
    • Google Scholar
42. 1. Li J
    2. Ji L

    (2005) Adjusting multiple testing in multilocus analyses using the eigenvalues of a correlation matrix

    Heredity 95:221–227.

    https://doi.org/10.1038/sj.hdy.6800717

    • PubMed
    • Google Scholar
43. 1. Likhtik E
    2. Stujenske JM
    3. Topiwala MA
    4. Harris AZ
    5. Gordon JA

    (2014) Prefrontal entrainment of amygdala activity signals safety in learned fear and innate anxiety

    Nature Neuroscience 17:106–113.

    https://doi.org/10.1038/nn.3582

    • PubMed
    • Google Scholar
44. 1. Lissek S
    2. Bradford DE
    3. Alvarez RP
    4. Burton P
    5. Espensen-Sturges T
    6. Reynolds RC
    7. Grillon C

    (2014) Neural substrates of classically conditioned fear-generalization in humans: a parametric fMRI study

    Social Cognitive and Affective Neuroscience 9:1134–1142.

    https://doi.org/10.1093/scan/nst096

    • PubMed
    • Google Scholar
45. 1. Martin A
    2. Rief W
    3. Klaiberg A
    4. Braehler E

    (2006) Validity of the brief patient health questionnaire mood scale (PHQ-9) in the general population

    General Hospital Psychiatry 28:71–77.

    https://doi.org/10.1016/j.genhosppsych.2005.07.003

    • PubMed
    • Google Scholar
46. 1. Meins E
    2. McCarthy-Jones S
    3. Fernyhough C
    4. Lewis G
    5. Bentall RP
    6. Alloy LB

    (2012) Assessing negative cognitive style: development and validation of a short-form version of the cognitive style questionnaire

    Personality and Individual Differences 52:581–585.

    https://doi.org/10.1016/j.paid.2011.11.026

    • PubMed
    • Google Scholar
47. 1. Mumford JA
    2. Poline JB
    3. Poldrack RA

    (2015) Orthogonalization of regressors in FMRI models

    PLoS One 10:e0126255.

    https://doi.org/10.1371/journal.pone.0126255

    • PubMed
    • Google Scholar
48. 1. Nili H
    2. Wingfield C
    3. Walther A
    4. Su L
    5. Marslen-Wilson W
    6. Kriegeskorte N

    (2014) A toolbox for representational similarity analysis

    PLoS Computational Biology 10:e1003553.

    https://doi.org/10.1371/journal.pcbi.1003553

    • PubMed
    • Google Scholar
49. 1. Onat S
    2. Büchel C

    (2015) The neuronal basis of fear generalization in humans

    Nature Neuroscience 18:1811–1818.

    https://doi.org/10.1038/nn.4166

    • PubMed
    • Google Scholar
50. 1. Palminteri S
    2. Justo D
    3. Jauffret C
    4. Pavlicek B
    5. Dauta A
    6. Delmaire C
    7. Czernecki V
    8. Karachi C
    9. Capelle L
    10. Durr A
    11. Pessiglione M

    (2012) Critical roles for anterior insula and dorsal striatum in punishment-based avoidance learning

    Neuron 76:998–1009.

    https://doi.org/10.1016/j.neuron.2012.10.017

    • PubMed
    • Google Scholar
51. 1. Pasternak T
    2. Greenlee MW

    (2005) Working memory in primate sensory systems

    Nature Reviews Neuroscience 6:97–107.

    https://doi.org/10.1038/nrn1603

    • PubMed
    • Google Scholar
52. 1. Patton JH
    2. Stanford MS
    3. Barratt ES

    (1995) Factor structure of the Barratt impulsiveness scale

    Journal of Clinical Psychology 51:768–774.

    https://doi.org/10.1002/1097-4679(199511)51:6<768::AID-JCLP2270510607>3.0.CO;2-1

    • PubMed
    • Google Scholar
53. 1. Pearson RM
    2. Heron J
    3. Button K
    4. Bentall RP
    5. Fernyhough C
    6. Mahedy L
    7. Bowes L
    8. Lewis G

    (2015) Cognitive styles and future depressed mood in early adulthood: the importance of global attributions

    Journal of Affective Disorders 171:60–67.

    https://doi.org/10.1016/j.jad.2014.08.057

    • PubMed
    • Google Scholar
54. 1. Resnik J
    2. Paz R

    (2015) Fear generalization in the primate amygdala

    Nature Neuroscience 18:188–190.

    https://doi.org/10.1038/nn.3900

    • PubMed
    • Google Scholar
55. 1. Resnik J
    2. Sobel N
    3. Paz R

    (2011) Auditory aversive learning increases discrimination thresholds

    Nature Neuroscience 14:791–796.

    https://doi.org/10.1038/nn.2802

    • PubMed
    • Google Scholar
56. 1. Rigoux L
    2. Stephan KE
    3. Friston KJ
    4. Daunizeau J

    (2014) Bayesian model selection for group studies - revisited

    NeuroImage 84:971–985.

    https://doi.org/10.1016/j.neuroimage.2013.08.065

    • PubMed
    • Google Scholar
57. 1. Rogan MT
    2. Leon KS
    3. Perez DL
    4. Kandel ER

    (2005) Distinct neural signatures for safety and danger in the amygdala and striatum of the mouse

    Neuron 46:309–320.

    https://doi.org/10.1016/j.neuron.2005.02.017

    • PubMed
    • Google Scholar
58. 1. Sasaki Y
    2. Nanez JE
    3. Watanabe T

    (2010) Advances in visual perceptual learning and plasticity

    Nature Reviews Neuroscience 11:53–60.

    https://doi.org/10.1038/nrn2737

    • PubMed
    • Google Scholar
59. 1. Schechtman E
    2. Laufer O
    3. Paz R

    (2010) Negative valence widens generalization of learning

    Journal of Neuroscience 30:10460–10464.

    https://doi.org/10.1523/JNEUROSCI.2377-10.2010

    • PubMed
    • Google Scholar
60. 1. Seymour B
    2. Daw ND
    3. Roiser JP
    4. Dayan P
    5. Dolan R

    (2012) Serotonin selectively modulates reward value in human decision-making

    Journal of Neuroscience 32:5833–5842.

    https://doi.org/10.1523/JNEUROSCI.0053-12.2012

    • PubMed
    • Google Scholar
61. 1. Seymour B
    2. O'Doherty JP
    3. Dayan P
    4. Koltzenburg M
    5. Jones AK
    6. Dolan RJ
    7. Friston KJ
    8. Frackowiak RS

    (2004) Temporal difference models describe higher-order learning in humans

    Nature 429:664–667.

    https://doi.org/10.1038/nature02581

    • PubMed
    • Google Scholar
62. 1. Shalev L
    2. Paz R
    3. Avidan G

    (2018) Visual aversive learning compromises sensory discrimination

    The Journal of Neuroscience 38:2766–2779.

    https://doi.org/10.1523/JNEUROSCI.0889-17.2017

    • PubMed
    • Google Scholar
63. 1. Slivinske AJ
    2. Hall JF

    (1960) The discriminability of tones used to test stimulus-generalization

    The American Journal of Psychology 73:581–586.

    https://doi.org/10.2307/1419947

    • Google Scholar
64. Book

    1. Spielberger CD
    2. Gorsuch RL
    3. Lushene RE

    (1970)

    The State-Trait Anxiety Inventory: Test Manual for Form X

    Palo Alto, CA: Consulting Psychologists Press.

    • Google Scholar
65. 1. Struyf D
    2. Zaman J
    3. Vervliet B
    4. Van Diest I

    (2015) Perceptual discrimination in fear generalization: mechanistic and clinical implications

    Neuroscience & Biobehavioral Reviews 59:201–207.

    https://doi.org/10.1016/j.neubiorev.2015.11.004

    • PubMed
    • Google Scholar
66. Book

    1. Sutton RS
    2. Barto AG

    (1998)

    Reinforcement Learning: An Introduction

    MIT Press.

    • Google Scholar
67. 1. Tibshirani R

    (1996)

    Regression Shrinkage and Selection via the Lasso

    Journal of the Royal Statistical Society. Series B  58:267–288.

    • Google Scholar
68. 1. Treanor M
    2. Barry TJ

    (2017) Treatment of avoidance behavior as an adjunct to exposure therapy: Insights from modern learning theory

    Behaviour Research and Therapy 96:30–36.

    https://doi.org/10.1016/j.brat.2017.04.009

    • PubMed
    • Google Scholar
69. 1. Tzourio-Mazoyer N
    2. Landeau B
    3. Papathanassiou D
    4. Crivello F
    5. Etard O
    6. Delcroix N
    7. Mazoyer B
    8. Joliot M

    (2002) Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain

    NeuroImage 15:273–289.

    https://doi.org/10.1006/nimg.2001.0978

    • PubMed
    • Google Scholar
70. 1. van Dam LC
    2. Ernst MO

    (2015) Mapping shape to visuomotor mapping: learning and generalisation of sensorimotor behaviour based on contextual information

    PLoS Computational Biology 11:e1004172.

    https://doi.org/10.1371/journal.pcbi.1004172

    • PubMed
    • Google Scholar
71. 1. Vlaeyen JW
    2. Linton SJ

    (2012) Fear-avoidance model of chronic musculoskeletal pain: 12 years on

    Pain 153:1144.

    https://doi.org/10.1016/j.pain.2011.12.009

    • PubMed
    • Google Scholar
72. 1. Walther A
    2. Nili H
    3. Ejaz N
    4. Alink A
    5. Kriegeskorte N
    6. Diedrichsen J

    (2016) Reliability of dissimilarity measures for multi-voxel pattern analysis

    NeuroImage 137:.

    https://doi.org/10.1016/j.neuroimage.2015.12.012

    • PubMed
    • Google Scholar
73. 1. Weinberger NM

    (2007) Associative representational plasticity in the auditory cortex: a synthesis of two disciplines

    Learning & Memory 14:1–16.

    https://doi.org/10.1101/lm.421807

    • PubMed
    • Google Scholar
74. 1. Wichmann FA
    2. Hill NJ

    (2001) The psychometric function: I. Fitting, sampling, and goodness of fit

    Perception & Psychophysics 63:1293–1313.

    https://doi.org/10.3758/BF03194544

    • PubMed
    • Google Scholar
75. 1. Wigestrand MB
    2. Schiff HC
    3. Fyhn M
    4. LeDoux JE
    5. Sears RM

    (2017) Primary auditory cortex regulates threat memory specificity

    Learning & Memory 24:55–58.

    https://doi.org/10.1101/lm.044362.116

    • PubMed
    • Google Scholar

Author details

1. Agnes Norbury

   Computational and Biological Learning Laboratory, Department of Engineering, University of Cambridge, Cambridge, United Kingdom

   Contribution

   Conceptualization, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing

   For correspondence

   aen31@cam.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-4377-3164

2. Trevor W Robbins

   1. Department of Psychology, University of Cambridge, Cambridge, United Kingdom
   2. Behavioural and Clinical Neuroscience Institute, University of Cambridge, Cambridge, United Kingdom

   Contribution

   Conceptualization, Methodology, Writing—review and editing

   Competing interests

   No competing interests declared

3. Ben Seymour

   1. Computational and Biological Learning Laboratory, Department of Engineering, University of Cambridge, Cambridge, United Kingdom
   2. Center for Information and Neural Networks, National Institute of Information and Communications Technology, Suita City, Japan

   Contribution

   Conceptualization, Supervision, Funding acquisition, Methodology, Writing—review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-1724-5832

• 6,562

  views

• 499

  downloads

• 31

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.