Objective This paper presents classifications of m ental tasks based on EEG signals using an adaptive Radial Basis Function (RBF) n etwork with optimal centers and widths for the Brain-Computer Interface (BCI) s che...Objective This paper presents classifications of m ental tasks based on EEG signals using an adaptive Radial Basis Function (RBF) n etwork with optimal centers and widths for the Brain-Computer Interface (BCI) s chemes. Methods Initial centers and widths of the network are s elected by a cluster estimation method based on the distribution of the training set. Using a conjugate gradient descent method, they are optimized during train ing phase according to a regularized error function considering the influence of their changes to output values. Results The optimizing process improves the performance of RBF network, and its best cognition rate of three t ask pairs over four subjects achieves 87.0%. Moreover, this network runs fast du e to the fewer hidden layer neurons. Conclusion The adaptive RB F network with optimal centers and widths has high recognition rate and runs fas t. It may be a promising classifier for on-line BCI scheme.展开更多
基金ThisworkwassupportedbytheNationalNaturalScienceFoundationofChina (No .3 0 3 70 3 95 )
文摘Objective This paper presents classifications of m ental tasks based on EEG signals using an adaptive Radial Basis Function (RBF) n etwork with optimal centers and widths for the Brain-Computer Interface (BCI) s chemes. Methods Initial centers and widths of the network are s elected by a cluster estimation method based on the distribution of the training set. Using a conjugate gradient descent method, they are optimized during train ing phase according to a regularized error function considering the influence of their changes to output values. Results The optimizing process improves the performance of RBF network, and its best cognition rate of three t ask pairs over four subjects achieves 87.0%. Moreover, this network runs fast du e to the fewer hidden layer neurons. Conclusion The adaptive RB F network with optimal centers and widths has high recognition rate and runs fas t. It may be a promising classifier for on-line BCI scheme.
