Thesis Topic: Meta-Learning of Reward Representations for Predicting Mood Fluctuations in Mental Health
A Master’s thesis topic is available at the Centre for Digital Health Interventions.
CDHI Thesis Title: Meta-Learning of Reward Representations for Predicting Mood Fluctuations in Mental Health: From Inverse Reinforcement Learning to Deep Sequence Models
Overview
We are seeking a motivated and hard-working Master’s student in computer science, engineering, or artificial intelligence with a strong background in machine learning and deep learning and a curiosity for applying these methods to digital mental health. This project offers the opportunity to contribute to computational psychiatry and make a real-world impact through AI-driven adaptive tools.
Project Purpose & Vision
In today’s fast-paced and unpredictable world, the ability to know when to persist, when to let go, and when to shift gears is key to sustaining agency and emotional balance. When this adaptability falters, rigid control patterns and maladaptive learning can emerge, contributing to psychological distress and mental health challenges.
The Gearshift Fellowship (GF) is a digital serious-game platform that measures and trains adaptability under shifting demands and uncertainty. It provides a mental mirror, a computational feedback system, that helps users understand how their behavior and beliefs shape motivation, emotion, and learning when facing stress and control loss. By integrating behavioral data with computational modeling, GF aims to probe and adjust the mechanisms linking adaptability and mental health.
Project Goals
This thesis will develop deep learning models that extract interpretable adaptability profiles from GF game behavior and neurocomputational markers. The overarching goal is to identify behavioral and computational digital signatures of mood and attention symptoms by modeling how individuals update their internal reward representations under uncertainty and stress. This project offers a unique opportunity to combine deep learning with computational psychiatry, contributing to the development of adaptive digital tools for mental health.
Key Tasks
The student will:
- Apply inverse reinforcement learning (IRL) to infer individual reward functions and controllability estimates from behavioral data collected in the GF environment.
- Implement disentangled recurrent neural network and/or transformer-based models to capture context-specific versus context-general adaptation patterns.
- Analyze whether these representations reveal meta-learning principles (i.e., how individuals adapt their learning strategies over time) and how such principles relate to mood fluctuations and mental-health outcomes.
Data & Resources
Behavioral (choices and response times) data and self-report questionnaire data will be provided from ongoing GF studies, along with code templates for data preprocessing and model implementation. Game pattern data include behavioral trajectories and decision logs from different game missions designed to probe neurocognitive constructs such as effort- and reward-sensitivity, cognitive control, and avoidance.
Required Skills
- Proficiency in Python and deep learning frameworks (PyTorch or TensorFlow)
- Practical experience with neural network architectures
- Familiarity with reinforcement learning
- Interest in computational cognitive modeling (a plus but not required)
Supervision
Supervisors: Dr. Nadja Ging-Jehli, Adaptive Intelligence & Mental Health Mechanisms Core, Center for Digital Health Interventions (CDHI), ETH Zurich, University of St. Gallen, University of Zurich;
Prof. Dr. Tobias Kowatsch, Center for Digital Health Interventions (CDHI), ETH Zurich, University of St. Gallen, University of Zurich
Contact
Nadja R. Ging-Jehli, PhD
Core Director of Adaptive Intelligence & Mental Health Mechanisms at the Centre for Digital Health Interventions, University of St. Gallen / ETH Zurich / University of Zurich
Independent Project Leader of Gearshift Fellowship