An overview of statistical learning theory
IEEE Transactions on Neural Networks
| Hi-index | 0.00 |
When using a pattern recognition technique to classify signals, a common practice is to define a set of features to be extracted and, possibly after feature selection/projection, to use a learning machine in the resulting feature space for classification. However it is not always easy to devise the "right" set of features for a given problem, and the resulting classifier might turn out to be suboptimal because of this, especially in presence of noise or incomplete knowledge of the phenomenon. In this paper we present an off-line comparison of three methods (genetic algorithm, time-delay neural network, support vector machines) that leverage different ideas to handle features; we apply them to the recognition of the P300 potential in an EEG-based brain-computer interface. They all performed good, with the genetic algorithm being slightly better.