My research is about developing self-supervised deep neural network models to denoise electromagnetic brain recordings when ground truth is not available..
Electromagnetic recordings are multi-sensor time-series data and inherently contaminated with noise.
The noise can be divided into two major sources - sensor noise which is independent for each sensor, and environmental noise which is correlated across the sensors.
The recorded data is therefore a mixture of brain activity (signal of interest), sensor noise and environmental noise and must be decomposed into its components. While commonly used frameworks are based on linear decomposition methods like SVD, PCA and ICA, non-linear methods as deep neural networks exceed their capability.