The mission of the Computational Cognitive Neuroscience group is to formulate quantitative theories of human brain function and to empirically validate these using non-invasive functional neuroimaging methods (EEG and fMRI). Our current work clusters in four project areas that are fundamentally related by the notions of probabilistic modelling and inference.
The ability to infer the dynamic statistical properties of the environment is a hallmark of cortical function and a necessary precondition for allostatic behaviour. In this project cluster, we investigate the degree to which these neural mechanisms can be described by probabilistic inference schemes. To this end, we combine behavioural experimentation with non-invasive brain imaging and computational modelling.
Over the last decade, model-based fMRI has successfully delineated a number of brain systems that are thought to constitute the neural substrate of habitual stimulus-response learning. However, it is generally accepted that human action-selection is only insufficiently described by immediate habitual associations from external states to actions, and, in many contexts, more realistically conceived as originating from a planning process that involves the anticipation, evaluation, and comparison of action-outcome contingencies. Such planning processes are fundamental to a class of decision problems which are referred to as “goal-directed” or "“sequential”. In this project cluster, we combine semi-naturalistic sequential decision paradigms across various domains to study their neural mechanisms.
We have previously used a set of information-theoretic functionals defined on a “model-free” probabilistic model to study the correlative structure of simultaneous EEG-fMRI and behavioural data features to arbitrary order. In this work, we could show that combining EEG and fMRI time-domain features by quantifying the information in their joint distribution is more informative about external stimuli than treating each modality in isolation. In more recent work, we use a “model-based” approach that embeds delay differential equation systems describing the interaction between cortical units in fully probabilistic forward models for EEG-fMRI. In brief, based on the assumption of high temporal (and low spatial) reliability of the EEG signal, and high spatial (and low temporal) reliability of the fMRI signal, the aim of this “Bayesian sensor fusion” approach is to enbable inference about spatiotemporal neural dynamics with both high spatial and high temporal confidence.
Variational Bayes is a deterministic-approximate inference framework for probabilitistic models. In the brain imaging field, variational Bayes enjoys widespread popularity as a general theory of brain function and as an inversion scheme for models of functional neural connectivity under the labels "Free Energy Principle" and "Dynamic Causal Modelling", respectively. In this project cluster, we explore the analytical underpinnings and qualitative properties of variational Bayesian methods.