Ahstract-Due to the non-stationarity nature and poor signal-to-noise ratio (SNR) of brain signals, repeated time consuming calibration is one of the biggest problems for today's brain-computer interfaces (BCls). In order to reduce calibration time, many transfer learning methods have been proposed to extract discriminative or stationary information from other subjects or prior sessions for target classification task. In this paper, we review the existing transfer learning methods used for BCI classification problems and organize them into three cases based on different transfer strategies. Besides, we list the datasets used in these BCI studies.

The classifications for movement-related potentials (MRPs) are used to provide control signals for many motor-related brain-computer interfaces (BCIs). A discriminative spatial pattern (DSP) algorithm has been shown to be effective for extracting MRPs. However, the spatial filtering and feature extraction of DSP are not exactly consistent. In addition, DSP filtering and subsequent classifiers, such as support vector machines, are optimized toward different goals. These two drawbacks may degrade overall classification performance. In this paper, inspired by the multilayer perceptron, we propose a hierarchical framework based on a logistic regression model. Our framework directly extracts the distance between each pair of time series as a feature, and unifies spatial filtering and classification under a regularized empirical risk minimization problem. Experimental results from three BCI datasets recorded from five subjects demonstrate that our method can make more accurate classifications.Keywords Movement-related potentials (MRPs) Á Braincomputer interfaces (BCIs) Á Discriminative spatial pattern (DSP) Á Multilayer perceptron (MLP)

No abstract

The famous common spatial patterns (CSP) algorithm has shown to be useful for event-related desynchronization (ERD) feature extraction of multi-channel electroencephalogram (EEG) signals. Actually, CSP only extracts the linear correlation between each pair of channels. The performance of CSP severely depends on the preprocessing. Moreover, CSP and the subsequent classifier are not optimized by the same criteria. In this paper, we investigated the nonlinear correlation between channels with kernel technique, and proposed a unified prediction framework based on linear ridge regression model. This prediction framework integrates preprocessing, feature extraction and classification, can automatically select the time windows, frequency bands and regularization parameter by minimizing leave-one-out crossvalidation error through gradient descent. Experimental results on the dataset IV, BCI competition II show the effectiveness of our approach.