Atypical Uncertainty Estimation Across Psychopathology: Insights from a Novel Learning Task

Student thesis: Doctoral ThesisDoctor of Philosophy


“We sail within a vast sphere, ever drifting in uncertainty, driven from end to end”.
B. Pascal, 1852

Background: To survive, humans and animals must adapt behavior in response to outcomes. However, a central question in cognitive science is how the brain learns from outcomes associated with different choices, given that outcomes can be highly uncertain. To account for the remarkable skill of optimal inference and decision making in the face of a capricious world, it is proposed that the properties of uncertainty itself must be accurately estimated. These properties include the changeability (or volatility) and noisiness of outcomes. Drawing on a normative account of learning, the joint estimation of noise and volatility is suggested to determine the speed of learning itself. Using (neuro) computational models, systematic deviation from normative learning has been linked to dysfunctional decision making in psychiatric illness. However, research to date has focused on properties of uncertainty in isolation i.e., noise or volatility. Although vital research, this approach fails to consider the full problem facing the learner: that learning involves the simultaneous estimation of both volatility and noise (i.e., stochasticity). Additionally, while aberrant uncertainty estimation has been linked to multiple psychiatric disorder categories, novel dimensional approaches to psychiatric nosology support the identification of commonly perturbed biobehavioral mechanisms as a route towards improved treatment. Computational psychiatry, the mechanistic description of maladaptive behaviour, used in conjunction with large-scale online populations, has emerged as a promising tool for capturing how cognitive deficits relate to the continuous, rather than binary (on/off; healthy vs. ill) nature of mental illness. As such, this thesis aimed to develop, test, and implement a novel online learning task that was able to capture learning about multiple properties of uncertainty. This task was then used to test whether aberrant uncertainty estimation could represent a fundamental feature across symptoms of psychopathology.

Methods: Firstly, a systematic review (patients/controls n=5371), assessed research that used (error-based) computational modelling to the question of belief updating in psychosis, depression, and anxiety disorders. Secondly, a novel learning task was developed. The task sampled from “ground truth” probability distributions that differed in their statistical configurations, designed to simulate different properties of uncertainty within the learning environment. The task included continuous response metrics that indicated the individuals’ estimate of both the variance (noise) and the mean on a trial wise basis. Two experiments tested the task on feasibility, reliability, and construct validity (n=555; n=167). Post validation, the task was applied to an eating disorder group (n=116) and healthy controls (n=67); a general population group with natural variation in across psychosis, depression and anxiety symptomology (n=580); and a selected population of high (n=81) and low (n=66) schizotypy traits. Generalised linear mixed effects models were used for learning data. An extended version of the Hierarchical Gaussian Filter (HGF: JGET) that “learns” about (1) the current mean value and (2) its variance of the input stream, was also applied (Chapter 6).

Results. The systematic review revealed both overlap and dissociation between disorder groups. For example, aberrant learning in response to the valence of feedback specifically was found in depression, and not anxiety disorder. Anxiety and psychosis disorders instead showed overlap regarding a difficulty in learning about, and adapting to, volatility. The novel task, built to specifications set out in the review, was demonstrated to be feasible in an online setting and displayed good levels of reliability and construct validity across two studies (n=555; n=167). Experimental evidence showed partial evidence of a valence dependent deficit in depression only. Elevated psychosis and anxiety traits, and eating disorder patients (vs. healthy controls), overlapped in relation to an inflexibility in adapting to increases in volatility. Elevated psychosis traits were additionally associated with overall non-normative learning rates in isolated conditions of noise and volatility. Computational modelling applied to data from elevated schizotypy groups revealed a tendency to incorporate their belief about the noise into decision making when noise was irrelevant.

Conclusion. This thesis demonstrates the ability for experimental design, and modelling approach, to explicitly test multiple dimensions of learning under uncertainty. In doing so, this research was able to provide insights into the simultaneous estimation of different aspects of uncertainty during learning. While further testing is required, a difficulty in adapting to increases in volatility may underlie multiple aspects of psychopathological symptomology, including eating disorder patients, and elevated trait anxiety and psychosis-related symptomology. However, elevated trait psychosis additionally showed a non-normative pattern of learning. In this case, it may be that difficulties in estimating the type of uncertainty e.g., mistaking noise for volatility, led to a trade-off effect in which estimates of volatility reached ceiling level. This could then lead to downstream impacts on both overall learning rates in volatility, and the ability to appropriately adapt to increases in volatility. Novel computational modelling approaches, for example the HGF: JGET were instrumental in capturing learning about the mean and variance independently and should be applied to future research interested in the joint estimation of uncertainty. Findings from the HGF: JGET suggest that schizotypy may be associated with an overloading of superfluous information when confronted with increased volatility, potentially due to a lack of confidence in a model of the world (i.e., volatility). Future research should build on these findings by investigating the interdependence of inferences about volatility and noise, and their relationships to unique and cross cutting symptoms of psychopathology.
Date of Award1 Jan 2024
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorKelly Diederen (Supervisor) & Thomas Spencer (Supervisor)

Cite this