A cognitive-computational model of avoidance learning

Learning to avoid threat and to seek safety is a fundamental psychological function. It allows us to flexibly adapt to ever-changing environments. This learning process is also leveraged in exposure therapy, a common clinical intervention for anxiety disorders.

Recent research has provided us with unprecedented detail on the underlying neuroanatomy, and with a range of novel candidate interventions for clinical conditions. However, there are two crucial shortcomings that impair both our theoretical understanding of aversive learning, and a clinical application. The first is a dearth of mechanistic models that predict behaviour outside of well-characterised experimental paradigms. As a consequence, even small procedural changes - unavoidable in many application settings - can have a major impact on the success of an intervention. The second is a focus on threat prediction, rather than avoidance of threat. This means that most research has used experimental paradigms that don't afford avoidance. Crucially, however, biological data suggests there are at least two partly independent learning systems for prediction and avoidance. Avoidance is arguably more relevant than prediction from a functional, biological, and clinical perspective. Hence, research into avoidance learning, and its relation to prediction learning is warranted.

This proposal seeks to fill both of these gaps, by developing a computational learning model that encompasses both threat prediction and avoidance. Using a virtual reality approach with combined threat prediction and threat avoidance measurement, we will first gather a large body of data (N = 800) in experimental paradigms that are diagnostic on the underlying learning systems. These data will be made publicly available for the community. We will then develop an array of computational learning models that can explain these data. To disambiguate between these models in further diagnostic experiments, we will leverage a Bayesian experimental design approach. Because behavioural data can be ambiguous, we will further provide neural evidence for the final learning model. To this end, we will use 7 T functional neuroimaging of brain stem regions important for neural signalling of learning quantities. Finally, we will benchmark the identified learning model by acquiring data in new experimental paradigms that were not used for model development. To provide direct application potential, these experiments will be designed with a view to maximising avoidance reduction in the model. Thus, the resulting experimental paradigms could be leveraged for development of clinical interventions.

This research proposal will provide computational learning models for threat prediction and avoidance that explain behaviour in unseen learning situations. This will crucially contribute to our psychological understanding of threat learning, and could thus form the cornerstone of an improved learning theory. As a powerful application, it will bring us a step closer to quantitative development of clinical interventions, and already has a potential to provide the blueprint of a clinical procedure. Finally, it will furnish a big and deep data set (overall N > 1000), which can be used by the academic community for development of computational models, theory, and measurement methods.