A Computational Model reveals Learning Dynamics during Interpretation Bias Training with Clinical Applications

Abstract

Background

Some psychopathologies, including those of anxiety and irritability, are associated with biases when judging ambiguous social stimuli. Interventions targeting these biases, or interpretation bias training, are amenable to computational modeling to describe their associative learning mechanisms. Here, we translate ALCOVE, a category learning modeling, to describe learning in youth with affective psychopathology when training on more positive judgments of ambiguous face emotions.

Methods

Seventy-one young people (ages 8-22 years) comprised a predominantly clinical sample to represent broad distributions of irritability and anxiety symptoms. Of these, sixty-three were included in the test sample, by completing an interpretation bias training task with acceptable performance for computational modeling. We used a separate sample of twenty-eight to translate ALCOVE for individual estimates of learning rate and generalization. In the test sample, we assessed associations between model learning estimates and irritability, anxiety, their shared variance (negative affectivity), and age.

Results

Age and affective symptoms were associated with category learning during interpretation bias training. Lower learning rates were associated with higher negative affectivity common to anxiety and irritability. Lower generalization, or improved discrimination between face emotions, was associated with increasing age.

Conclusions

This work demonstrates a functional consequence of age- and symptom-related learning during interpretation bias. ALCOVE measured learning also revealed learning types not accounted for in the prior literature on interpretation bias training. The work more broadly demonstrates the utility of measurement models for understanding trial-by-trial processes and identifying individual learning styles.

Introduction

People with affective psychopathology demonstrate biased social judgments that correspond with their predominant mood (e.g, negative interpretations of ambiguous social information). Cognitive theories suggest these biases maintain a pathologic state by influencing downstream cognition towards pathology-congruent behaviors (e.g., ^1–3). Following these theories, interpretation biases have become a target of diverse training programs aimed at improving affective psychopathology. These training programs often encourage more positive expectations from ambiguous stimuli through reinforcement learning. Thus far, clinical trials of these programs have had mixed outcomes, including our own efforts to target hostile interpretation bias in youth with severe irritability or anger (4,5). Such mixed results suggest a need to better understand mechanism underlying learning processes during interpretation bias training (6–8). Here, we apply a computational model representing core principles of category learning theory to measure learning during interpretation bias training in a clinical sample.

Interpretation Bias Training (IBT) is a relatively new form of cognitive bias modification with a procedure that is amenable to a close examination of its mechanism. When forced to choose “happy” or “angry”, individuals with pathologic irritability (4) or anxiety (9) tend to interpret ambiguous facial expressions as angry. During IBT, individuals are encouraged to change their emotion valence label of an ambiguous facial expression from negative to positive (e.g., angry to happy; Figure 1). Categorical training of emotional judgments on ambiguous face emotions results in stable changes in classification of these face emotion stimuli over weeks (4, 10) and generalizes to new face emotion stimuli. Such a shift towards more positive judgments of ambiguous faces should be related to symptom improvement (7, 11).

Figure 1.

An interpretation bias training session. First an indifference point is assessed, where a person’s judgments of morphs switches from predominantly happy to angry. Training sessions provide feedback on judgments that encourage happy judgments of two ambiguous faces previously judged as angry. Training effects are measured by another indifference point assessment.

At least two factors have been identified that might influence learning during IBT in youth with affective psychopathology: The ability to learn new response associations targeting a bias towards hostile interpretations of ambiguous facial expression (4) and the ability to discriminate between differences in facial expressions (12). Of note, this ability improves as children grow to adulthood (13, 14). These two processes- the ability to flexibly change valence associations with feedback and to discriminate between differing facial expressions- likely contribute to learning ability during IBT. They may be potential treatment targets or explain variability in patient response to treatment. To examine these contributions to training effects, a first step is to measure them as people learn during IBT.

Learning during IBT may be characterized with validated computational models which are expressions of categorical learning theory. The models represent components of learning including error-driven changes to stimulus-response associations (15) and the similarity structure of exemplar-based, categorical representations across stimuli (16). These major components may be integrated in a single model to examine associative learning for stimuli which may have membership in multiple categories (17). The model, as an expression of category learning theory, allows us to examine trial-by-trial behavior. It yields core metrics of individual performance. Model-based metrics include: 1) the learning rate of updating new stimulus response associations during training and 2) generalization, the degree to which learning associations to one stimulus affects neighboring stimuli associations.

The learning rate will reflect an individual’s ability to flexibly change categorical associations during IBT. Generalization reflects the ability to discriminate between stimuli on a continuum between one overt facial expression to another, with higher values suggesting less discrimination between neighboring stimuli. Together these model-based measures indicate how an individual learns during training. Though this work is focused on interpretation bias training, it joins emerging efforts to measure learning dynamics during behavioral interventions using computational models. Notable examples are models designed for functional communication training (e.g., ¹⁸) and exposure therapy (e.g., ¹⁹). Overall, these efforts translate models grounded in learning theory to measure learning dynamics within an intervention’s training sessions. Such models provide important clues to mechanism of action for the intervention and provide indicators of learning ability.

Modeling learning during training distinguishes this work from complementary efforts that associate an intervention response to learning during assessments that are separate from the intervention (e.g., ²⁰). With regards to interpretation bias on face emotions specifically, it also differs from conventional approaches of generating summary statistics across task trials (e.g. (21–23, 14). Though these may be used to measure a pre-post intervention change, this approach is limited in its representation of trial-wise dynamics, such as speed at which a person acquires new associations during IBT. For example, it may be that individual differences in learning rate or generalization are important features of psychopathology. In addition, from a treatment perspective, measuring learning during a training session will provide important information on how an individual learns. This may help in intervention design, such as determining “dosing,” e.g. the amount of training trials people typically require to shift their indifference point, or to develop adaptive training based on an individual’s learning ability. Learning rate may be an important indicator of treatment response. For instance, the ability of a person to make associations during operant learning tasks is a predictor of treatment response in adult depression (e.g., ^24–25). Indeed, in preliminary application of the proposed model to measure a form of IBT targeting bipolar depression, unadjusted learning rate during the very first session was a predictor of treatment response (26).

The study represents a model-based examination of the mechanism of learning during IBT by testing its associations with anxiety and irritability in youth. Via model parameters, we test associations between affective psychopathology and two learning indicators, generalization and learning rate. Based on our pilot work, we expected a negative association between learning rate and the negative affectivity common to irritability and anxiety. We also expected a positive association between generalization and negative affectivity, as higher generalization reflects worse discrimination. Finally, given that face-emotion discrimination improves with age, we expected a negative association between age and generalization.

Beyond symptom associations, we expect that the model may be used to help identify how individuals learn during IBT, which may provide insights into the mechanism of action of IBT, predict treatment outcomes and differentiate approaches to the task that are not currently considered in trials. Following the experimental therapeutics framework, this work should elucidate mechanisms of action of IBT for its revision as well as better identify individuals who might benefit from it. Ultimately, this work represents a translation of cognitive modeling for the application of understanding and treating affective psychopathology.

Methods

Participant Characteristics

Participants were 71 youth (8-22 years old). Of these, 17 had no major psychopathology. The other 54 had been diagnosed with psychopathology characterized by irritability (Disruptive Mood Dysregulation Disorder), anxiety (any of Generalized, Social or Separation Anxiety Disorders) or a relevant comparison condition, Attention Deficit/Hyperactivity Disorder. All were participating in ongoing research in the Emotion and Development Branch at the National Institute of Mental Health, with data collected November 2015 through December 2018. Diagnoses were determined by KSADS-PL by trained clinician interview with specific module for DMDD (27–28). Briefly, exclusion criteria were major medical illness, substance use, major depressive disorder, psychotic disorder, intelligence quotient (IQ) estimated at <70, and autism spectrum disorder. The rationale for accruing on these criteria is to generate a sample of youth with a full joint distribution of clinically relevant anxiety and irritability symptoms (e.g., ²⁹).

Participant characteristics for those providing acceptable data (n=63) are reported in Table 1. They completed the 6-month Affective Reactivity Index (ARI); (30) parent-report and youth-report forms to assess levels of chronic irritability and the Screen for Child Anxiety Related Emotional Disorders (SCARED) (31) parent-report and youth report forms to assess levels of anxiety within three months of their visit. Most completed forms during a visit to the NIMH, on average 6.6 to 8.1 days, depending on the measure, prior to completing the IBT task.

Table 1.

Participant Characteristics^a

Primary Diagnosisb
DMDD (n)	21
ADHD (n)	16
ANX (n)	9
HV (n)	17
Female (n, %)	28, 44%
Age (M, SD) years	14.68, 3.23
IQ (n, M, SD)	51, 116.65, 10.63
Irritability (ARI)
Self-report (n, M, SD)	53, 2.87, 2.65
Parent-report (n, M, SD)	53, 3.47, 3.47
Anxiety (SCARED)
Self-Report (n, M, SD)	53, 17.20, 14.18
Parent-Report (n, M, SD)	53, 13.44, 14.13

^a.DMDD = Disruptive Mood Dysregulation Disorder; ADHD = Attention Deficit/Hyperactivity Disorder; ANX = any of Generalized Anxiety Disorder, Social Anxiety Disorder, or Separation Anxiety Disorder; HV = no major psychopathology.

^b.Individuals with a primary diagnosis may overlap with others without such diagnoses.

Interpretation Bias Training Task

Participants completed one training session of IBT. The IBT task (Figure 1) required participants to make happy-angry judgments of 15 face-emotion stimuli of a composite male face from the Karolinska Directed Emotional Faces (32). Ambiguous face-emotion stimuli were created by morphing the composite happy and angry faces on an equally spaced continuum (Figure 1); these types of stimuli are often called “morphs”.

The task was comprised of 8 blocks. In the first, participants rated 3 randomly presented replicates of each morph with no feedback. The task then roughly estimated a person’s indifference point by the proportion of the number of happy judgments over the number of all responses multiplied by 15 (33). Feasibly calculated by the presentation software (Eprime 2.0), this yields a point along the morph continuum (ordered overly happy to angry) close to the indifference point, which is defined as the morph at which the probability of an angry judgment is 0.5 (4, 26).

In blocks 2 through 7, participants judged 12 randomly presented replicates of each morph. They received feedback with a training threshold equal to the estimated indifference point plus two morphs towards angry rounded down to the nearest integer. The morph continuum was divided at the feedback threshold. For morphs on the happy side and at the feedback threshold, participants received positive feedback for happy responses and negative feedback for angry responses. Feedback was reversed on the angry side of the morph continuum (Figure 1). In the final block, participants repeated the first block for post training ratings of faces without feedback.

All trials began with a fixation cross (1500-2000 ms), morph presentation (150 ms), visual noise mask (250 ms), and response screen with a question mark until response. For training trials, feedback (“Right!/Wrong! That face was happy/angry.”) occurred after response for 1500 ms.

Computational Model

We model error-driven learning during forced judgments of facial stimuli with ambiguous representations of emotional expression in a connectionist framework (17). There are some differences from ALCOVE. The pretraining block in IBT allowed for an initialization of expected valence association weight, i.e., the pretraining expectation of happy/angry categories. Two pretraining bias parameters were added to account for asymmetric biases individuals exhibit when judging facial affect (21). When these parameters are equal, they represent the chance of guessing, thus they are called ‘guessing parameters’.

The model is parameterized such that learning rate and generalization are both considered together to estimate a maximum effective learning rate. Changes in valence judgments of a stimulus on the morph continuum are influenced by learning from neighboring stimuli, which is determined by generalization. Thus, learned changes in valence differs by any point, i,in the stimulus continuum and is represented by the effective learning rate ${{\epsilon }^{\mathrm{eff}}}_i$. This relationship is a consequence of ALCOVE; it is derived and specified in supplemental information. Learning rate is reported here in two ways. First as it is fit, an estimate of the maximum generalization-adjusted effective learning rate, ${{\epsilon }^{\mathrm{eff}}}_{\max }$, which is limited to an upper bound of 1. Second, given the focus of the task and analysis, the effective learning rate about the training threshold is calculated, ${{\epsilon }^{\mathrm{eff}}}_{\mathrm{thr}}$.

Model architecture is depicted in Figure 2 and its fitted parameters are described in Table 2. Its build, equations, derivation of effective learning rate, fitting, and variant testing in an independent sample are detailed in supplemental information. The model was fit to the pretraining and training blocks 1-7 only. The model’s expected indifference point at the end of training block 7, p(Angry Judgment) = 0.5, was compared to an indifference point estimated by a 4-parameter logistic curve (21) fit to the post training categorical identification block data, block 8. This last block was not fit to the computational model. For parameter recovery tests, one thousand synthetic datasets were generated by random selection of parameters from a uniform distribution ranging from their mean +/− 2 standard deviation values from the current sample. Input parameters were correlated to parameters estimated by model fit to synthetic data.

Figure 2.

Model architecture. A person’s judgment of the emotional valence of a face depends on their expectation of the face’s emotional valence (v). In turn, this depends on the perception of emotional features in the face, (a). This is shared with similar face stimuli, the degree to which it is shared is given by σ. The expected valence also depends on prior valence associations to perceived emotion in faces (w). Association weights update according to feedback encoded in prediction error (d) proportional to the learning rate (e) and to emotion activation (a). Red: Free parameters of interest, estimated from behavioral data. Blue: Internal variables, with learning dynamics predicted by the model.

Table 2.

Estimated Parameters

Parameter	Role	Interpretation	Mean, Median N=63	Parameter Recovery Correlation
Initial weight matrix indifference point (p)	Estimate expected valence of each morph before learning.	Point on the morph scale at which the weights for both response options are equal.	7.63, 7.89	0.66
Initial weight matrix slope (s)		Change in weight towards an angry response for arbitrary weight units per morph over the morph range. Thus, s scales the initial weight matrix before training.	0.30, 0.18	0.59
Guessing (gA, gH)	Model other sources of variance, e.g. noise, between expected valence and response.	Error rate for extreme stimuli for the angry (gA) and happy (gH) of the morph continuum. Their sum is equal to the overall error rate; their difference represents a choice bias in error. gH – gA is an angry choice bias index on overt facial expressions, with positive values representing excess angry judgments on overt facial expressions.	gA 0.060,0.038 gH 0.053,0.018	0.77 0.88
Inverse temperature (θ)		Determines the probability of an angry choice from the expected valence for any morph. For very high values of θ, the probability of an angry judgment is either 0 or 1, depending on the expected valence of angry. For very low values of θ, the probability of an angry choice is ½ for all morphs, regardless of expected valence.	4.49, 2.72	0.45
Generalization (σ)	Model stimulus perception, representation, and learning.	The degree to which feedback on one morph affects the updating of neighboring morph activations. σ is the standard deviation in morph units of a Gaussian centered on any morph.	6.73, 5.04	0.76
Learning rate (εeffmax)		Determines the degree to which prediction error contributes to updating the valence association weights.	0.21, 0.10	0.76

^a.Divergence between mean and median values reflects a bimodal distribution in the patient group for θ and s, due to individuals with near deterministic responses, or a long tail for guessing parameters and effective epsilon.

Analysis

Given that these youth are known to have impairments in generally engaging in tasks, participants were only included if they demonstrated adequate engagement. Prior to analysis, we applied two arbitrary performance criteria based on the post-training assessment data to which the model is not fit. The first is that individuals have at least 75% accuracy on labeling four overt facial expressions (n=5 with unacceptable accuracy). The rationale is to ensure attention throughout the training session. The second is that an individual’s final indifference point is 2 morphs lower than their starting indifference point (i.e., 4 morphs lower than the training threshold in the wrong direction, n=3). The rationale is that cases where such a large difference occurs generally reflect degrading performance or noncompliance for which the model has no mechanism. Sixty-three of 71 youth had acceptable data.

Analyses were completed in R version 4.1.0, Matlab 2019b and MPlus version 8. To indicate irritability or anxiety, we used a latent model of irritability and anxiety initially validated by Kircanski et al. (34). This model allows for a weighted account of individual items from the parent and self-report ARI and subscales of SCARED, reporter effects, and addresses collinearity in raw scores by partialling out shared variance. To promote stable individual estimates, the bifactor model was fit to raw ARI items and SCARED subscale scores from a super sample which included participants in this study and others with the same inclusion criteria concurrently accrued (N=353) (35). It had acceptable fit statistics (RMSEA=0.075, TLI=0.965, SRMR=0.064). The fit bifactor model provided estimated factor scores for symptom dimension: shared variance (negative affectivity), parent-reported irritability, self-reported irritability, and combined parent- and youth-reported anxiety. Of the 63 individuals with acceptable IBT data, 54 had provided enough symptom measure responses to generate symptom dimension scores following Cardinale etal., 2019 (35). Of the 54, 4 were each missing one measure.

We entered these factor scores along with those for the general factor representing shared variance between irritability and anxiety, negative affectivity, into two univariate linear regressions as additive predictors. The dependent variables were the two model-based measures of interest: effective learning rate about the training threshold and generalization. For these regressions, dependent variables were Tukey transformed to improve normality, with λ=0.325 for ${{\epsilon }^{\mathrm{eff}}}_{\mathrm{thr}}$ and 0.4 for σ.

We empirically identified categorical learning types through mixture modeling on model-based measures of learning. Learning type, also known as class in mixture modeling, was identified through the Mplus version 8 Mixture argument to the Analysis command with a maximum likelihood estimator with robust standard errors. This classification of participants into different types of learners was done on the following model-based measures: 1) the effective learning rate at the training threshold, ${{\epsilon }^{\mathrm{eff}}}_{\mathrm{thr}}$; 2) generalization, σ; 3) model estimated indifference point just prior to the training blocks; and 4) a bias towards angry judgments on overt facial expressions. The latter two measures are of broad interest and calculated from the model. As noted above, the indifference point reflects a person’s angry/happy response bias for ambiguous stimuli and may be queried from the model at any point during the task; higher values reflect a bias towards happy judgments. The response bias on overt faces can be calculated from the guessing parameters described above. In this constrained, 2-alternative forced choice task, gH represents excess angry judgments of overtly happy faces and gA represents excess happy judgments for overly angry faces. An inequality between the two represents a response bias, with higher values reflecting an angry response bias. For tests of group differences between these classes, t-tests for single variables or logistic regression multiple variables were used.

Associations between continuous variables, such as age and generalization, were by correlation tests, Pearson’s for parametric and Kendall’s tau for nonparametric. Analytic and model code is posted at www.github.com/joelStod/FaceEmo.

Results

Parameter Characteristics

Parameter descriptions, distribution, as well as estimation precision and recovery for this sample are presented in Table 2. There is good parameter recovery for each of the two main parameters of theoretical interest (generalization, σ, and effective learning rate, ${{\epsilon }^{\mathrm{eff}}}_{\max }$; each r = 0.76) as well as separation between them (r = −0.04). The computational model predicted future behavior as indicated by an association between post-training indifference point predicted by the model and an indifference point estimated from post-training face-emotion categorization block that was not input to the model r(60) = 0.62, p < 0.01.

Learning Types and Affective Psychopathology

There was an expected, negative association between age and generalization in the full sample r(61) = −0.32, p = 0.01. No association between age and learning rate was detected τ(61) = −0.10, p = 0.23. Regression of learning rate on symptom dimensions (Table 3) demonstrated a significant negative association between learning rate and shared variance between anxiety and irritability (negative affectivity) but a positive association to unique variance for each of anxiety and parent report irritability. No association was detected between generalization on symptom dimensions in regression; adding age as a covariate did not change this result.

Table 3.

Associations between Psychopathology and Model-based Measures

	Generalization (σ)	Effective Learning Rate about the Training Threshold (${{\epsilon }^{\mathrm{eff}}}_{\mathrm{thr}}$)
Intercept	2.0 (0.11)***	0.49 (0.037)***
General (Negative Affectivity)	0.15 (0.16)	−0.11 (0.052)*
Self-report Irritability	0.20 (0.18)	0.13 (0.060)*
Parent-report Irritability	−0.07 (0.20)	0.14 (0.068)*
Anxiety	0.17 (0.16)	0.16 (0.053)**
R2	0.11	0.18
F(4,49)	1.6	2.8*

Parameter estimates (b) and standard error (se), formatted b (se), in two univariate regressions predicting generalization or learning rate about an individual participant’s training threshold. The predictors are latent variable scores across symptom measures of anxiety and irritability representing their common variance (general factor or negative affectivity), irritability by reporter, and the common report of anxiety.

^†p<0.1,

^*p<.05,

^**p <.01,

^***p<.001

There were distinctive types of learners during IBT, with a best fit gaussian mixture modeling on learning rate, generalization and response bias empirically classifying individuals into two distinctive learning types (Fig. 3; BIC=518.9, entropy=0.948, Lo-Mendell-Rubin likelihood ratio test p=.0003, see supplemental information for model fit and selection table). The first group (n=20) was characterized by high generalization, M(SD) σ= 13.88 (1.99), low learning rates, M(SD) ε^eff_thr=0.14 (0.14), and a tendency to judge ambiguous faces as happy, M(SD) pretraining Indifference point = 8.2 (1.6)]. The second group (n=43) was characterized by variable learning with lower generalization [n=43, M(SD) σ= 3.40 (2.29), M(SD) ${{\epsilon }^{\mathrm{eff}}}_{\mathrm{thr}}$=0.22 (0.28), M(SD) pretraining indifference point =7.1 (1.7)]. For those in the first category, a model-free plot of the data demonstrates the influence of prior trial effects of distant stimuli on ambiguous faces at the training threshold (Fig. 4). As expected from the association between generalization and age, those in the high generalization group were younger M(SD) age=13.3(2.6) versus 15.3(3.3) years, t(47.6)=2.67, p=0.011. No associations between learning groups and dimensions of psychopathology were detected either in individual t-tests (unadjusted p’s > 0.14) or logistic regression with age as a covariate (p’s>0.29).

Figure 3.

Learning typology is evident in the model parameters. Individuals clearly differ in the degree to which they are affected by feedback on the prior trial (generalization). The arbitrary, empirical classification suggests lower, less variable learning rates (top) and higher indifference points (bottom) in the group characterized by high generalization.

Figure 4.

A model free illustration of high generalization by prior trial feedback. Generalization refers to the influence of neighboring stimuli, or morphs, on learning. Its effects are clearly represented without the need for a model in a single ambiguous morph. This morph is highly vulnerable to uncertainty because it is the morph just on the happy side of the feedback threshold during training (see Figure 1). Recall, the position of the training threshold on the morph continuum differs for each participant by their initial balance point so, the morph just on the happy side of this threshold is denoted as morph 0 (the gray line) for reference. To represent the distance on the morph continuum of prior trials, the judgments of morph 0 are divided into categories that are defined by its distance to the morph in the prior trial. Relative to morph 0, prior trial morphs may be distant (4 to 7 morphs to the left of the feedback threshold) or may be close (within 3 morphs of either side of the threshold). Feedback during the prior trial encourages happy judgments on the left side of the gray line, and it encourages angry judgments to the right. High generalizers are clearly influenced by prior trial feedback, even if the feedback was to a morph distant to morph 0 (p<.01). Note, this represents 1/15^th of the data, i.e. judgments of morph 0, and ‘n’ refers to the number of judgments of morph 0 across participants.

Discussion

This study investigated associative learning during IBT, a form of cognitive bias modification training, and its relationship to affective psychopathology in youth. Negative affective symptoms are associated with the learning rate of positive judgments to ambiguous facial stimuli. With generalization and learning rate, we found distinctive learning types which are associated with age. Notably, this work converges with an emerging literature demonstrating that generalization in face emotion category learning is a function of age. This report makes several contributions to the literature. By translating a category learning model for its application to interpretation bias training, it provides a way to embed analysis and interpretation of treatment development of IBT in a robust category learning theory. The translation of the model and associations found here have implications for determining the mechanism of IBT as well as identifying individuals who might benefit from it.

ALCOVE translated to IBT well, making it possible to measure learning during IBT in a principled fashion. However, the association between model parameters to symptoms was less robust than expected. There was an expected negative association between learning rate and negative affectivity, as reliable symptom dimension derived from measures of irritability and anxiety from different informants (35, 36). Interestingly, once a shared variance of negative affectivity is partialled out, anxiety and irritability were positively associated with learning rate. These dimensions are less reliable than negative affectivity, so their associations should be interpreted with caution. Speculatively they may indicate the effects of an increase in anxiety or irritability specific increase in attention to angry features of facial expression that may enhance learning once negative affectivity is accounted for. Regardless, it is important to appreciate that our understanding of these transdiagnostic dimensions of psychopathology is in its infancy.

Contrary to our expectations, we did not detect an association between generalization and symptoms, even when accounting for age. We did detect a decrease in generalization with age which is consistent with the finding that adolescents are increasingly able to perceptually discriminate between face-emotion stimuli with age (14). However, this is to the extent that generalization during learning is determined only by perceptual confusion between stimuli. Another consideration is that age influences the ability of individuals to make distinctive cognitive inferences between perceptually discriminable items (37). The latter consideration is consistent with evidence that categorical identification relies on a broadly distributed set of visual and semantic neural regions that interact to resolve ambiguity (38, 39), even for social and affective representations (40). Overall, this work converges with prior work to suggest that developmental effects on generalization are an important consideration for interpretation bias training.

A striking finding of this work is that by measuring learning dynamics we distinguished types of learners during IBT. About a third of the sample learned in a way that is not consistent with IBT’s intended manipulation. That is, they learned to positively respond along large portion of the continuum of stimuli as opposed to IBT’s intended mechanism of targeting a change in responses to ambiguous stimuli only. Different approaches to learning have not been considered in prior trials. It is important to note here that we applied standard exclusion criteria to avoid including participants who were simply judging overtly angry faces as happy. One interpretation is that the broad generalizers are susceptible to “demand effects”, i.e. they are complying with a belief that the experimenter is requesting generally positive responses. However, the association generalization with age suggests that there may be a more specific, developmental limitation in discriminating between face stimuli that leads to an alternative approach to the task apart from demand effects. Differences in learning styles suggest that these may need to be accounted for in recruitment and analysis. Indeed, generalization itself may be an alternative treatment target, specifically improving an individual’s precision in emotional categorization and learning.

Limitations of this experiment are the modestly sized sample of children and adolescents who have known difficulty with general task engagement. In addition, these participants completed an fMRI task using these stimuli a few days prior to this experiment which may have affected engagement. Taken together, this study may be vulnerable to type II error in detecting associations between model-based measures of learning and psychopathology. Still, such data provided a stress test of model fits which were reasonably predictive of future behavior and allowed for model optimization in a sample representing the clinical population of interest. Another limitation of the current model build is the requirement of learning to segregate cognitive processes to determine expected valence apart from other sources affecting response probabilities modeled by the choice rule. However, integrating reaction time into modeling choice data will improve both this model of associative learning as well as other models of simple categorical identification. Finally, this work only assesses concurrent clinical associations and not the associations between learning dynamics and clinical response to IBT, which is an important direction for future work.

Conclusion

The current model allows an investigation of the mechanism of interpretation bias training from a category learning framework, incorporating a generalization model of the relationship between morphs. It is applicable to current trials of interpretation bias training, but also more broadly to cognitive bias modification, and investigations into basic social information processing impairments in individuals with affective psychopathology.