Table 8.

Summary of classification methods.

	Approach	Properties	Applications
Generative model	Bayesian analysis	– Assigns the observed feature vector to the labeled class to which it has the highest probability of belonging – Produces nonlinear decision boundaries – Not very popular in the BCI systems	[245–248]
Linear	LDA	– Simple classifier with acceptable accuracy – Low computation requirements – Fails in the presence of outliers or strong noise. Regularization required – Usually two class. Extended multiclass version exits. – Improved LDA versions: BLDA, FLDA	[179,230,231,233–235]
	SVM	– Linear and non-linear (Gaussian) modalities – Binary or multiclass method – Maximizes the distance between the nearest training samples and the hyperplanes – Fails in the presence of outliers or strong noise. Regularization required – Speedy classifier	[131,228,230, 237,239–244]
Non-linear
	k-NNC	– Uses metric distances between the test feature and their neighbors – Multiclass – Efficient with low dimensional feature vectors. Very sensitive to the dimensionality of the feature vectors	[227–229]
	ANN	– Very flexible classifier – Multiclass – Multiple architectures (PNN, Fuzzy ARTMAP ANN, FIRNN, PeGNC)	[200,215, 249–256]