Joram Soch: Cross-validated Bayesian model selection

 

BCCN Berlin

In cognitive neuroscience, functional magnetic resonance imaging (fMRI) data are routinely
analyzed using general linear models (GLMs) without systematic model quality control. This
can lead to underfitting of the data by using a model that is too simple or overfitting of the
data by using a model that is too complex. We have recently developed cross-validated
Bayesian model selection (cvBMS), a technique that identifies the optimal model from a set
of candidate models, given data from a group of subjects (Soch et al., NeuroImage, 2016).
This method is based on the cross-validated log model evidence (cvLME), a Bayesian model
quality criterion for which no prior distributions need to be specified.
In the theoretical part of my talk, I will introduce the basic concepts of classical and Bayesian
model selection, give a brief overview on fMRI data analysis using the GLM and present an
application of cvBMS to fMRI. In the practical part of my talk, I will introduce PycvBMS, a
Python module that allows to calculate the cvLME for GLMs ( https://github.com/
JoramSoch/PycvBMS ), and demonstrate an application to spike count data obtained with
Neuropixel probes (analyzed in the tutorial of this seminar). You should all come, because
proper model quality control increases the replicability of your results.

 

Organized by

Joram Keijser, Filip Vercruysse, Henning Sprekeler

Location

BCCN Berlin, lecture hall, Philippstr. 13 Haus 6, 10115 Berlin