Our trainees review webinars in their given fields and share abstracts to help colleagues outside their discipline make an informed choice about watching them. As our program bridges diverse disciplines, these abstracts are beneficial for our own group in helping one another gain key knowledge in each other’s fields. We are happy to share these here for anyone else who may find them helpful.
Dynamic causal modelling: Tutorial and first results for multi-brain data
Edda Bilek
March 18, 2020
Transcultural Psychiatry
Analysis by Olawale Salaudeen:
Dynamic Causal Modeling (DCM) is a framework for learning complex dynamics from data. This framework uses differential equations to describe the interactions between neural populations that give rise to the blood-oxygen-level-dependent (BOLD) signal captured by functional magnetic resonance imaging (fMRI) and other functional neuroimaging methods.
This tutorial gives a good overview of DCM that can be instructional for those with basic knowledge of neuroscience and neuroscientists. The speaker describes the mapping from the DCM equations to brain-specific motivations, i.e., connecting terms in the DCM equations to state changes, connectivity, modulation of connectivity, system states, direct inputs, and external inputs. They also discuss important technical topics such as the Bilinear state equation, Bayesian parameter estimation, and Bayesian model reduction in a broadly accessible way.
DCM is a type of causal learning that imposes strong structural assumptions — motivated by domain expertise. Unlike some more generic causal learning paradigms that have to discover causal structure along with the functional form of the generative mechanisms, DCMs aim to learn parameters that best explain the observed data, given the structure of the Bilinear state equation. One important difference between DCMs and other general causal learning models is that there is no Acyclicity assumption, in fact, bilinear state equation directly models acyclicity. This is consistent with the belief that neural states have self-feedback loops. For example, Linear Non-Gaussian Acyclic Models also assume linearity but do not allow for acyclicity.