Between 2013 and 2015 I investigated and developed a variety of machine learning approaches for real world applications of EEG devices.
ICA and Dynamic Information
Independent Component Analysis (ICA) is a common preprocessing tool used in the analysis of EEG data. Potentials measured at the scalp at various locations are highly correlated, but each represents a different mixing of the sources of electrical activity being measured. Under the assumption of linear, instantaneous mixing, ICA seeks 1-dimensional linear subspaces of the signal space that are maximally statistically independent. This is possible because, generally, independently generated sources that are randomly mixed result in component distributions (in this case, the signals recorded at each channel) that are more Gaussian.
I have developed a novel approach to Blind Source Separation (BSS) that extends ICA to incorporate dynamic, or rate-of-change, information. My approach is called Pairwise Complex ICA (PWC-ICA); the primary paper detailing the approach was published in Computational Intelligence in 2016, and source code is available. A presentation demonstrating the basic approach and some of our results on simulated data is available below.
Manifold Regularization and Transfer Learning
I spent some time investigating the method of manifold regularization as a transfer learning approach that might be useful for some of the classification tasks we face in our lab. Manifold regularization is an extension of RLS and SVM classifiers that, as I understand it, modifies the classification hyper-surface (and hence the separating submanifold that specifies a a classifier) over the data space through a soft constraint involving both labeled and unlabeled data.
I’ve included a presentation I delivered on the topic below: