Probabilistic brain decoding priors for regional activation and non-stationary spatial correlation

A key problem in functional magnetic resonance imaging
(fMRI) is to predict behavioral variables from massive and
high-dimensional brain activation patterns, referred to as "brain
decoding", which remains a difficult challenge. Most previous analyses
disregard some critical principles of brain activities: regional
activation and non-stationary spatial correlation. To deal with the
high dimensionality of fMRI data, we regularize the learning model
with the above-mentioned structural principles so that the extracted
features are more informative and interpretable. In this talk, I will
present my work on developing two regularizing priors for
probabilistic brain decoding problems.

Firstly, I will talk about a prior that encourages region-sparse
activation for brain decoding. In many problem settings, weight
vectors are not merely sparse but dependent in such a way that
non-zero coefficients tend to cluster together. We refer to this form
of dependency as "region sparsity." We introduce a Gaussian process
based hierarchical model for such region-sparse weight vectors, which
models parameters as dependent a priori. We show that the proposed
model provides spatial decoding weights for brain imaging data that
are both more interpretable and achieve higher decoding performance.

Secondly, I will talk about a brain-decoding prior that encodes
non-stationary spatial correlation structure across brain. Standard
smoothing methods ignore non-stationary spatial correlations, such as
discontinuities across boundaries and bilateral symmetries. But such
non-stationarity can be estimated from functional correlations of
voxels. To cope with that, we introduce the "brain kernel", a
continuous covariance function for full-brain activity patterns that
captures shared intrinsic functional anatomy of the brain across
individuals estimated from massive resting-state fMRI (rfMRI) data. We
show that by incorporating functional correlation extracted from
rfMRI, the brain kernel can be used as a prior covariance function for
probabilistic modeling for brain decoding and factor analyses on four
task fMRI datasets with improved performance. 

BIO : Anqi Wu is a Postdoctoral Research Fellow at the Center for
Theoretical Neuroscience at the Zuckerman Mind Brain Behavior
Institute at Columbia University, working with Prof. Liam Paninski and
Prof. John Cunningham. She received her Ph.D. degree in computational
and quantitative neuroscience with Professor Jonathan Pillow from
Princeton University in 2019. She holds bachelor's and Master's
degrees in electrical and computer science from Harbin Institute of
Technology in China and University of Southern California. Her
research interests lie broadly in Bayesian statistical models of
high-dimensional and large-scale neural response and fMRI decoding.
She is especially interested in sparse Bayesian structure learning for
fMRI decoding with hierarchical generative models and Bayesian
nonparametric latent variable models for large-scale neural
recordings.