University of Cambridge > > Microsoft Research Machine Learning and Perception Seminars > Margin- and Evidence-Based Approaches for EEG Signal Classification

Margin- and Evidence-Based Approaches for EEG Signal Classification

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Oliver Williams.

Our principal motivation in Brain-Computer Interface (BCI) research is to develop systems that will enable a completely paralysed person to communicate. Many projects have adopted the approach of classifying changes in the bandpower in certain parts of the spectrum of brain signals measured by electroencephalogram (EEG). Appropriate preprocessing, in the form of spatial and temporal filtering, is crucial for this, and generally the `conventional wisdom` of the BCI community is (a) that once the preprocessing is right, it doesn`t matter which classifier you use and accordingly (b) that a linear classifier will generally perform as well as any non-linear one. The hitherto most successful preprocessing algorithm, the Common Spatial Pattern algorithm, is a supervised method for computing spatial filters very cheaply. However, it uses a least-square criterion and is very prone to overfitting with small amounts of data. Our approach is to combine feature extraction and classification into a single optimization step, and to optimize a criterion that is a better predictor of generalization performance: namely the margin (as in the Support Vector Machine) or the evidence (a.k.a. marginal likelihood, in this case obtained from a Gaussian Process classifier). I will show that this yields consistent improvements in performance, particularly in the (most clinically relevant) cases where data are noisy and/or few in number, and that projection into a higher-dimensional feature space via a non-linear kernel can improve performance further. I will also show a preliminary demonstration that the approach can simultaneously recover optimal weightings across space, frequency and time, with little sensitivity to prior assumptions: this makes it a promising tool for the analysis of biosignal data in general.

This talk is part of the Microsoft Research Machine Learning and Perception Seminars series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2019, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity