Temporal independent component analysis for separating event-related potentials

A pervasive problem in neuroscience is to determine which regions of the brain are active, given electroencephalographic (EEG) recordings at the scalp. Analysis of evoked potentials in EEG recordings is important approach to reveal the relation between brain functions and structures. In this paper, we present a new approach for evoked potentials extraction and localization, suggesting to explore the high order statistics and temporal structures of the evoked potentials. Efficient learning algorithm is developed for training the demixing model and temporal filters. Theoretic analysis and computer simulations are given to show performance and efficiency of the proposed temporal independent component analysis.