[Paper Review] Improving brain computer interface performance by data augmentation with conditional Deep Convolutional Generative Adversarial Networks

Qiqi Zhang, Ying Liu|arXiv (Cornell University)|Jun 19, 2018

EEG and Brain-Computer Interfaces20 references59 citations

TL;DR

This paper proposes a conditional DCGAN (cDCGAN) to generate artificial EEG data for data augmentation, aiming to improve CNN-based BCI performance on motor imagery with limited training data. The generated data can match or enhance testing accuracy and bolster classification when data are scarce.

ABSTRACT

One of the big restrictions in brain computer interface field is the very limited training samples, it is difficult to build a reliable and usable system with such limited data. Inspired by generative adversarial networks, we propose a conditional Deep Convolutional Generative Adversarial (cDCGAN) Networks method to generate more artificial EEG signal automatically for data augmentation to improve the performance of convolutional neural networks in brain computer interface field and overcome the small training dataset problems. We evaluate the proposed cDCGAN method on BCI competition dataset of motor imagery. The results show that the generated artificial EEG data from Gaussian noise can learn the features from raw EEG data and has no less than the classification accuracy of raw EEG data in the testing dataset. Also by using generated artificial data can effectively improve classification accuracy at the same model with limited training data.

Motivation & Objective

• Address the challenge of very limited training samples in brain-computer interfaces.
• Introduce a conditional Deep Convolutional GAN to generate artificial EEG data for augmentation.
• Assess whether GAN-generated data can learn meaningful EEG features.
• Evaluate the impact of augmented data on CNN-based BCI performance using motor imagery data.

Proposed method

• Adopt a conditional Deep Convolutional Generative Adversarial Network (cDCGAN) to produce artificial EEG signals.
• Generate artificial EEG data from Gaussian noise conditioned on class labels.
• Use the generated data to augment training sets for CNN-based BCI classifiers.
• Evaluate performance on the BCI Competition Dataset for motor imagery.
• Compare classification accuracy using real data alone vs. augmented datasets.
• Demonstrate that generated data can learn features from raw EEG and achieve comparable testing accuracy.

Experimental results

Research questions

• RQ1 Can cDCGAN-generated artificial EEG data improve CNN-based BCI classification when training data are limited?
• RQ2 Do the generated EEG samples learn features representative of real EEG data and preserve discriminative information?
• RQ3 Is augmentation with cDCGAN data able to match or exceed the accuracy achieved with real data alone in testing?

Key findings

• Generated artificial EEG data from Gaussian noise can learn features from raw EEG data.
• The generated data achieve no less classification accuracy than the raw EEG data in the testing dataset.
• Using generated artificial data can effectively improve classification accuracy when training data are limited.
• The approach is evaluated on the BCI competition dataset of motor imagery.
• The cDCGAN-based augmentation addresses small-sample challenges in BCI with CNN classifiers.