Electromagnetic Brain Imaging using Sparse Bayesian Learning – Noise Learning and Model Selection
share
Brain source reconstruction from electro- or magnetoencephalographic (EEG/MEG) data is an ill-posed inverse problem. In this setting, it is crucial to distinguish signal components in the data (originating from actual brain sources) from noise components that are due to measurement interference as well as artifacts. Reliable noise estimates can be obtained from baseline data. However, in many scenarios such as resting-state analyses, no such baseline exists, and it is common practise to use rules of thumb or heuristics instead. In this work, we present novel principled techniques to a) learn the noise variance from the data within the source reconstruction procedure, and b) estimate the noise variance from hold-out data.
READ FULL TEXT