Electrical & Computer Engineering     |     Carnegie Mellon

Tuesday, May 06, 12:15-1:15 p.m. HH-1112


Jinyin Zhang
Carnegie Mellon University

A Doubly Regularized Support Vector Machine for Automatic Channel Selection of Brain Computer Interface

In a real-time brain-computer interface (BCI), it is desirable to minimize the number of recording channels so that the system complexity (e.g., electrode number, data rate, communication bandwidth, signal processing hardware, etc.) can be reduced. While such a channel selection is a critical component to enable real-time BCI, it has not been extensively studied in both communities of neuroscience and machine learning. In particular, most algorithms developed for feature selection are not applicable here, because one channel is typically mapped to multiple features and, hence, simply minimizing the number of features does not necessarily result in the minimal number of channels. In this work, we propose a novel doubly regularized support vector machine (DrSVM) to address the aforementioned problem. Our objective is to achieve a high-quality classifier for movement decoding, while simultaneously minimizing the number of channels that are used. The proposed DrSVM is derived from the theory of sparse classifiers (in particular, L1-norm SVM) and is carefully tuned for BCI application. Most importantly, we develop the concept of doubly regularized SVM by introducing a set of additional variables to model the "importance" of every channel. These extra variables are subject to sparsity constraints so that solving the proposed DrSVM results in a minimal number of important channels. In addition, while the DrSVM formulation cannot be mapped to a convex programming, we propose a robust relaxation algorithm to find the optimal model coefficients. Such a relaxation algorithm consists of a sequence of linear programming steps and guarantees to converge to a local optimal point. The proposed DrSVM was tested by several experiments using Magnetoencephalography (MEG) of human subjects. Our preliminary results show that compared with the traditional L1-norm SVM, the proposed DrSVM reduces the number of selected channels by 5~20x, while achieving the same accuracy for classification.


Jinyin Zhang received her B.S and M.S degrees in Computer Science from Beijing University of Astronautics and Aeronautics in 2002 and 2005, respectively. From 2005 to 2007, she worked for Lucent Technologies China as an engineer in base station software development. She is currently working towards the Ph.D. in the department of Electrical and Computer Engineering at Carnegie Mellon University.