[Paper Review] EEG-GAN: Generative adversarial networks for electroencephalograhic (EEG) brain signals

Kay Gregor Hartmann, Robin Tibor Schirrmeister+1 more|arXiv (Cornell University)|Jun 5, 2018

Neural dynamics and brain function145 citations

TL;DR

The paper adapts a Wasserstein GAN with gradient penalty to generate single-channel EEG signals, explores architectural and training choices for time-series data, and evaluates several metrics to assess realism and diversity of generated EEG.

ABSTRACT

Generative adversarial networks (GANs) are recently highly successful in generative applications involving images and start being applied to time series data. Here we describe EEG-GAN as a framework to generate electroencephalographic (EEG) brain signals. We introduce a modification to the improved training of Wasserstein GANs to stabilize training and investigate a range of architectural choices critical for time series generation (most notably up- and down-sampling). For evaluation we consider and compare different metrics such as Inception score, Frechet inception distance and sliced Wasserstein distance, together showing that our EEG-GAN framework generated naturalistic EEG examples. It thus opens up a range of new generative application scenarios in the neuroscientific and neurological context, such as data augmentation in brain-computer interfacing tasks, EEG super-sampling, or restoration of corrupted data segments. The possibility to generate signals of a certain class and/or with specific properties may also open a new avenue for research into the underlying structure of brain signals.

Motivation & Objective

• Motivate generation of artificial EEG signals using GANs to enable data augmentation and analysis.
• Stabilize GAN training for time-series EEG data via improved Wasserstein GAN with gradient penalty.
• Investigate architectural choices for time-series generation (up-/down-sampling, convolutional sizes) to maximize realism.
• Compare multiple evaluation metrics to assess similarity between real and generated EEG signals.
• Discuss potential applications such as EEG augmentation, super-sampling, and data restoration.

Proposed method

• Adopted improved Wasserstein GAN training with a one-sided gradient penalty scaled by critic difference to stabilize training.
• Implemented progressive growing of GANs by increasing time-series resolution from 24 to 768 samples across 6 steps.
• Employed convolutional generator and critic architectures with upsampling/ downsampling blocks and 9-sized kernels, using leaky ReLUs and pixel normalization.
• Applied regularization techniques from progressive GAN literature (minibatch stddev, equalized learning rate, pixel norm).
• Evaluated generated EEG with multiple metrics including Inception Score (IS) using a pretrained EEG classifier, Frechet Inception Distance (FID), Euclidean distance (ED min), and Sliced Wasserstein Distance (SWD).
• Compared up-/down-sampling schemes (AVG pooling vs conv downsampling; nearest-neighbor vs linear vs cubic upsampling).

Experimental results

Research questions

• RQ1 Can GANs generate realistic single-channel EEG signals comparable to real data?
• RQ2 How do architectural and training choices (sampling, kernel size, up-/down-sampling) affect EEG signal generation quality?
• RQ3 Which evaluation metrics reliably reflect EEG signal realism and diversity?
• RQ4 Is training stable for EEG generation using GANs, and under what conditions does mode collapse occur?
• RQ5 What are potential applications of EEG-GAN outputs in neuroscience and clinical contexts?

Key findings

• A modified WGAN-GP can stabilize EEG signal generation and avoid collapse for certain architectures.
• Different architectures perform best under different metrics; no single metric encapsulates quality.
• IS is not a reliable indicator of EEG generative quality, while SWD and EDmin provide complementary insight; FID does not always align with spatial/spectral realism.
• The best overall spatial-spectral realism varied by metric, with CONV-LIN often performing well on IS and SWD, while CONV-CUB performed poorly on several metrics.
• WGAN-GP with their one-sided gradient penalty and distance-scaled regularization showed stable gradients and avoided collapse in their experiments (unlike the original WGAN-GP in one setup).
• The most natural-looking distributions were not always those with the lowest FID, highlighting metric limitations for EEG data.