Application of Autoencoders for Feature Extraction in BCI-SSVEP

A brain-computer interface (BCI) based on the steady-state visually evoked potentials (SSVEP) paradigm deals with the challenge of determining the frequency associated with the visual stimulus the user is concentrated on, given electroencephalography (EEG) recordings of the brain activity. For this, the BCI process the brain signals in order to remove artifacts and, more importantly, to extract relevant features that contribute to the classification. A technique known as autoencoder (AE) has gained special attention in the last decades due to its ability to discover advantageous representations for a dataset, even with a significant dimensionality reduction. Essentially, autoencoders (AEs) are neural networks composed of two parts—encoder and decoder—whose roles are, respectively, to create the internal representation (named code) for the input data, and to reconstruct the input data from the generated code. Thus, the encoder corresponds to a powerful nonlinear feature extractor. In this work, we investigated the use of AEs to perform the feature extraction in a BCI-SSVEP. Different AE approaches have been analyzed, both in time and frequency domains, considering two classifiers: logistic regression and support-vector machines. The obtained results reveal that AEs can offer a performance improvement when compared with a BCI using the discrete-time Fourier transform (DFT) features.