A regularized discriminative framework for EEG analysis with application to brain–computer interface

A Robust and Efficient Spatio-Temporal Feature Selection for Interpretation of EEG Single Trials

“…Palaniappan et al [8] and Schroder et al [9] found appropriate channels via a Genetic Algorithm; Lal et al [10] used Recursive Feature Elimination and Zero-Norm Optimization to reduce the number of electrodes from 39 to 12. Tomika and Muller [11] reduced the dimension of the data by down-sampling the signals. Another way to alleviate the curse of dimensionality is via regularization methods, which stabilize the solutions by introducing prior knowledge or by restricting the solution ( [12]- [13]).…”

“…Another way to alleviate the curse of dimensionality is via regularization methods, which stabilize the solutions by introducing prior knowledge or by restricting the solution ( [12]- [13]). Cross-validation can be used to find the optimized model and its regularization parameter ( [11], [14]- [16]). …”

“…Tomioka and Muller [11] suggested an EEG single trials spatio-temporal interpretation that was based on three different regularizers. The regularizers were used to reveal different and complementary aspects of the localization of the discriminative information.…”

“…Furthermore, reducing the sampling rate may ignore important properties of the signal, which are visible in the EEG high temporal resolution. The use of different regularizers [11] may be problematic as it may produce contrasting interpretations with no clear ability to determine which of them is more accurate.…”

“…This paper extends previous work [20] and includes a comparison to time-points-selection via ERP method and to the sequential fusion of time points method. This two-step framework leads to an effective feature selection using a small number of parameters, dozens compared with thousands in [11], while maintaining a high temporal resolution.…”