Adaptive learning under uncertainty and its relation to symptoms of mood disorders

Thesis

A key mechanism that may contribute to symptoms of anxiety and depression is how individuals learn under uncertainty. Understanding this process could inform the development of more eective treatments. This thesis focuses on two common types of uncertainty encountered in everyday decision-making: volatility, which involves systematic changes in the environment over time, and noise, which refers to random fluctuations in outcomes. These types of uncertainty demand opposite learning strategies: volatility requires rapid updating, while noise calls for slower learning to avoid overfitting. In addition, the thesis investigates momentum, a form of structured volatility characterized by gradual, directional change. While prior research suggests that individuals with higher anxiety symptoms show impaired adaptation to volatility, the underlying mechanisms remain unclear. One possibility is that they fail to detect volatility; another is that they confuse it with noise, leading to maladaptive learning.

To address these questions, this thesis presents a series of three studies. In the first study, we focused on developing a rigorous experimental and analytical framework. Specifically, we designed and validated a task in which both forms of uncertainty were independently manipulated and established a set of criteria for selecting outcome schedules. To characterize how learning adapted to these uncertainties, we compared model-free and model-based approaches for estimating learning rates. In the second study, we showed that human learners increased their learning rate with high volatility and reduced their learning rate with high noise. Individuals reporting greater anticipatory pleasure exhibited less adaptation to noise. Pupillometry revealed that higher anxiety symptoms correlated with reduced pupil response to volatility, with tonic pupil area tracking volatility only in low-anxiety participants. In the third study, we observed that momentum increased phasic pupil dilation, while noise increased tonic pupil diameter. Behaviorally, participants showed higher learning rates for prediction errors in the same direction as the momentum, demonstrating adaptation. We extended the Rescorla-Wagner model to account for momentum eects, finding significantly higher model-estimated momentum when momentum was present.

These findings demonstrate how human learners adapt both behaviorally and physiologically to dierent types of uncertainty, with individual dierences in mood disorder symptoms associated with variations in these adaptations. Future research should establish causal links between maladaptive learning and mood disorders to inform more eective interventions.

Authors: Shang, S

• DOI: 10.5287/ora-qmvzrjnm7
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: volatility; uncertainty; noise; mood disorder; computational psychiatry
• Subjects: Psychiatry