The characterization and recognition of electrical signatures of brain activity constitutes a real challenge. Applications such as Brain-Computer Interfaces (BCI) are based on the accurate identification of mental processes in order to control external devices. Traditionally, classification of brain activity patterns relies on the assumption that the neurological phenomena that characterize mental states is continuously present in the signal. However, recent evidence shows that some mental processes are better characterized by episodic activity that is not necessarily synchronized with external stimuli. In this paper, we present a method for classification of mental states based on the detection of this episodic activity. Instead of performing classification on all available data, the proposed method identifies informative samples based on the class sample distribution in a projected canonical feature space. Classification results are compared to traditional methods using both artificial data and real EEG recordings.