Studying dynamic functional connectivity can lead to better understanding of the brain. Modeling dynamic functional connectivity is met with challenges of high dimensionality and noisy data. The common models used in neuroimaging studies such as principal component analysis and hidden Markov models are either non-probabilistic or inflexible. We present a probabilistic model to learn low-dimensional latent connectivity dynamics with Gaussian processes.
Source code for the LFGP model:
factor_gp.py LFGP model class
inference.py Gibbs sampling algorithm for posterior inference
blr.py Bayesian linear regression with conjugate prior for factor loadings
metropolis.py Metropolis random walk (used within Gibbs) for GP hyper-parameters
mvn.py Multivariate Gaussian conditional distribution and covariance decomposition
High resolution plots
example.ipynb Fitting an example model and plotting the results
simulation.ipynb Simulating time series data with latent dynamic covariance for model comparison