“Correlated noise and its implications for functional connectivity estimation and machine learning model interpretation”
Hosted by the WashU Neuroimaging Community (WUNIC)
Abstract: Multi-channel neuroimaging data, including magneto- and encephalography (M/EEG), often display a rich correlation structure. These correlations can be leveraged by multivariate machine learning methods (also known as decoding or MVPA) to improve predictions by maximizing signal-to-noise ratio. However, the correct interpretation of multivariate models is hindered by noise correlations, which can lead to wrong conclusion about the brain. We illustrate this problem using simple examples and propose a remedy for linear methods. Moreover, we emphasize the implications for non-linear methods and demonstrate that so-called “explainable artificial intelligence” methods can lead to severe misinterpretations. In the second part of the talk, we discuss the influence of correlations on M/EEG functional connectivity (FC) estimates. Unlike in fMRI, instantaneous correlations in M/EEG data do not serve as proxy for brain interactions as they can trivially arise from electrical volume conduction in the head. Instead, robust FC metrics that require non-zero time-delays are needed. We present robust metrics to quantify directed and non-directed brain interactions, and we highlight their possible use as informative markers of brain function.
Bio: Stefan Haufe is joint Associate Professor of Uncertainty, Inverse Modeling and Machine Learning at Technische Universität Berlin – a German research university, and Physikalisch-Technische Bundesanstalt Berlin – the German National Institute for Metrology, as well as a Group Leader at Charité – Universitätsmedizin Berlin, Germany’s largest university hospital. His group develops signal processing, inverse modeling and machine learning methods primarily for analyzing neuroimaging data. He is also interested in quality assurance for machine learning including model interpretation and validation of “explainable artificial intelligence”. He holds a diploma in computer science from Martin-Luther Universität Halle-Wittenberg and a Ph.D. in machine learning from TU Berlin. Before starting his lab in Berlin, he was a postdoctoral fellow in the BME departments at the City College of New York and Columbia University.
For inquiries contact Cathy Gezella.