Use of Generative Adversarial Networks in Brain-Computer Interfaces Systems
Brain-computer interface (BCI) systems have been object of high interest due to their possible applications in medical and entertainment areas. For those tasks, usually, classification is performed based on signals that carry information about the cerebral activity. However, these signals have high complexity and are subject to artifacts, aggravated by inaccuracies in the measurement instrument. Deep neural networks have been recently used as an efficient tool to interpret the signal without the necessity of manually adjusting certain parameters. In this context, Convolutional Neural Networks show very interesting results in selective attention and motor imagery BCI paradigms, surpassing the accuracy obtained from traditional classification methods. On the other hand, these new solutions suffer due to the lack of available training data. This project aims to use Generative Adversarial Networks (GANs) to generate artificial signals to improve the training of the convolutional network. A mapping of the brain signals into images shall be used to achieve this objective, since these networks show more potential in dealing with multidimensional data. The use of transfer learning in GANs shall also be employed, to transform the data from a subject into a target subject. With those approaches, we aim at achieving more efficient BCI systems and a better understanding of brain signals.