Abstract

The Linear Dynamical Systems (LDS) model can be viewed as a type of dynamic factor analysis in which the factors are fixed but the factor loadings evolve according to a dynamical system. Maximum Likelihood model parameters can be estimated using Expectation-Maximisation (EM) or gradient-based methods, and a Variational Bayes (VB ) approach has been developed that allows for inference over parameters and model dimension (e.g. how many factors). In this talk I’ll described the application of VB-LDS to cognitive, white and gray matter data from Cam-CAN following the example of de Mooij et al, JoN, 2018. I’ll describe how we’re updating the EM/VB-LDS algorithms to accommodate data sets with observations that are completely or partially missing or which have multiple observations at the same time point (age). I’ll then propose that this methodology be used to integrate cognitive neuroscience data across multiple databases and discuss with you the challenges involved.