Online regularized discriminant analysis

Learning the signal statistics and calibration are essential procedures for supervised machine learning algorithms. For some applications, e.g ERP based brain computer interfaces, it might be important to reduce the duration of the calibration, especially for the ones requiring frequent training of the classifiers. However simply decreasing the number of calibration samples would decrease the performance of the algorithm if the algorithm suffers from curse of dimensionality or low signal to noise ratio. As a remedy, we propose estimating the performance of the algorithm during the calibration in an online manner, which would allow us to terminate the calibration session if required. Consequently, early termination means a reduction in time spent. In this paper, we present an updating algorithm for regularized discriminant analysis (RDA) to modify the classifier using the new supervised data collected. The proposed procedure considerably reduces the time required for updating the RDA classifiers compared to recalibrating them, that would make the performance estimation applicable in real time.