Synthetic Feature Generative Adversarial Network for Motor Imagery Classification: Create Feature from Sampled Data

Motor imagery (MI) classification based on deep learning (DL) has made great progress. These methods heavily rely on sufficient training samples, which makes a great challenge for MI classification due to the difficulty of collecting brain signals. Generative adversarial network (GAN) has been an efficient tool to augment data. However, it is difficult to generate EEG samples from Gaussian noise by existing GANs. Also, the two-phase processing (augmentation and classification) in GANs causes computational inefficiency. In addition, EEG signal carries important private information which must be protected appropriately and has been disregarded in previous studies. To address these issues, Synthetic Feature Generative Adversarial Network (SF-GAN) is proposed combining data augmentation and classification. The input Gaussian noise of the generator is replaced by EEG samples and the generator generates features instead of raw data. A classifier interacting with the generator is incorporated. The performance of SF-GAN was verified by extensive experiments. It has been observed that the distribution of the synthetic features covers the distribution of real features which can be directly used as the augment input to the classifier instead of raw data and well protect data privacy.