BCI & Data Acquisition

BCI systems are engineered to link brain activity with external devices for a multitude of applications. Currently, electroencephalography remains the most prevalent method for acquiring brain data in BCIs due to its non-invasive, portable, and cost-effective nature. Most projects necessitate substantial amounts of EEG data.

Electroencephalography (EEG) is defined as "an electrophysiological monitoring method to record the electrical activity of the brain." For non-invasive BCI applications, EEG is the primary method for capturing the brain's electrical activity.

However, the EEG method faces several limitations:

• The requirement for lengthy calibration sessions for participants.
• Potential data corruption during collection.
• Classifiers trained on EEG datasets often fail to generalize across different recording times, even for the same individual.
• EEG signals exhibit significant variability between individuals and across sessions.

These factors create challenges for real-time analysis, particularly when models trained on historical neural data are used to interpret current neural signals. Moreover, acquiring the necessary number of samples is often both difficult and costly. The lengthy calibration times associated with BCI systems remain a significant barrier to commercial viability.

A major limitation in the BCI sector is the scarcity of training samples, making it hard to develop a reliable system. Even transfer learning techniques struggle to classify different EEG signal types without sufficient subject-specific data.

Transfer learning is defined as "a supervised learning technique that reuses parts of a previously trained model on a new network tasked with a different but similar problem."

To overcome these challenges, there is a pressing need for a method to automatically generate synthetic EEG trials to augment existing datasets. Generative methodologies present a promising avenue to tackle these limitations.

By utilizing generative techniques, the robustness of classifiers can be enhanced in a cost-effective and time-efficient manner. The incorporation of synthetically generated EEG signals can significantly improve the accuracy and generalizability of models designed to classify EEG data.

Section 1.1: Data Augmentation and BCI

Two primary approaches exist for generating augmented data:

1. Applying geometric transformations such as translations, rotations, cropping, flipping, and scaling.
2. Introducing noise to existing training data.

By expanding the training dataset, we can enable the training of more complex models and reduce the likelihood of overfitting.

However, EEG data, being a collection of noisy, non-stationary time-series signals from various electrodes, poses unique challenges. Unlike images, where labelers can easily ascertain the similarity of augmented datasets to the original class, EEG signals complicate this process.

Data Augmentation Techniques for EEG

Data augmentation serves to enhance the available training data, allowing for the utilization of more intricate deep learning models. It can also mitigate overfitting, bolstering classifier accuracy and stability.

Data augmentation has seen varying levels of success in numerous EEG applications, such as:

• Sleep stage classification
• Motor imagery
• Mental workload analysis
• Emotion recognition tasks

Both GANs and Variational Auto-Encoders (VAEs) have proven effective in capturing essential features from diverse datasets to generate realistic samples.

Another strategy involves "generating artificial EEG trials from the limited EEG trials initially available to augment the training set size effectively." This process entails segmenting existing training EEG trials and creating new artificial trials by concatenating segments from different training trials.

Additional EEG data augmentation methods include (but are not limited to):

• Sliding window technique (a method to reformulate time series data as a supervised learning problem)
• Noise addition
• Sampling methods
• Fourier transformation (a technique for expressing periodic signals in terms of their frequency components)

Focus on GANs

GANs represent a novel machine learning approach wherein two neural networks train against each other, fostering competition. This mechanism allows machines to generate original images.

GANs and related deep learning algorithms have primarily been applied to create synthetic images in computer vision. EEG signals can be analyzed and visualized over time in the frequency domain as spectrograms (utilizing Fourier or wavelet transformations).

Spectrograms can then be treated similarly to images, permitting the application of GAN-derived data augmentation methods. The spectrograms produced through this DA process can subsequently be converted back into EEG signals.

Recent studies have attempted to apply GANs to EEG signals:

In one investigation, "EEG is recorded while a person observes specific images, and this EEG is then processed through GANs to recreate the displayed image." They found that GANs outperformed VAEs, although the generated images exhibited limited quality.

Another study applied GANs to enhance the spatial resolution of EEG signals. Additionally, a separate research effort utilized GANs to produce synthetic EEG data for emotion recognition by generating EEG in the form of differential entropy from a noise distribution using a conditional WGAN with two emotion recognition datasets.

Issues with Standard GANs

Standard GANs can be susceptible to errors and instability, with "mode collapse" being a notable issue. This occurs when the discriminator learns to recognize only a few narrow modes of the input distribution as authentic, resulting in a generator that produces a limited range of outputs.

Training GANs can be complex due to pervasive instability. One solution to this challenge is the adoption of Wasserstein GANs (WGAN), introduced by Arjovsky et al. The goal of WGANs is to minimize the discrepancy between real and generated data distributions, thereby preventing mode collapse. This methodology has been successfully employed to generate noiseless EEG data.

Other researchers have highlighted problems such as "vanishing/exploding gradients," which WGANs can mitigate using a gradient penalty. Some have opted for pure Deep Convolutional GANs (DCGAN) to capture signal features using convolutional layers with varying kernel sizes.