Bayesian modelling of learning and failure mechanisms in neurofeedback training

Overview & Synopsis

Brain-Computer Interfaces (BCIs) are attracting a great deal of interest from researchers and the general public alike. Many BCIs, particularly Neurofeedback (NF) systems, require users to learn how to regulate specific subsets of their own brain activity via a simple online feedback signal. BCIs and NF hold great promise, especially for therapeutic applications. However, their lack of reliability means that their widespread adoption is currently limited. This unreliability stems from poorly understood, multi-factorial causes. Significant sources of variability include: the uncertain mapping between mental strategies and observable neural biomarkers; the hidden nature of the user's mental states and learning processes; as well as inaccuracies and ambiguities in task design and feedback. These factors hinder both the interpretation of experimental outcomes and the user's ability to learn effective self-regulation strategies. Consequently, there is also a lack of consensus regarding what is learnt during BCI training, how this learning occurs and which experimental factors critically affect training outcome. To address these gaps, this thesis builds a quantitative understanding of BCI learning using Bayesian models through complementary lines of investigation: 1) I developed a formal mathematical description of the BCI loop, casting perception, decision-making and learning as Bayesian (Active) Inference; 2) I leveraged computational simulations of this model to understand its theoretical properties and generate hypotheses about factors limiting performance; 3) I designed and ran online behavioural tasks to isolate and test core assumptions about human learning under the types of uncertainty inherent in BCI/NF tasks; 4) I synthesized these components and fitted the developed models to real BCI training data, thereby illustrating how this principled approach could be used to explain observed learning patterns and identify sources of inter-individual variability. This work provides the NF and BCI research communities with a set of computational tools that pave the way to a formal, objective and quantitative characterization of individual learning trajectories. One of its major prospects is to be used to optimize these trajectories through tailor-made adaptation throughout the training path.