Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving SSVEP BCI Spellers With Data Augmentation and Language Models

Joseph Zhang, Ruiming Zhang, Kipngeno Koech, David Hill, Kateryna Shapovalenko

TL;DR

This work tackles the variability and generalization challenges of SSVEP-based BCI spellers by coupling data augmentation with a language-model-based prior. The authors establish a strong EEGNet baseline on the Benchmark dataset and systematically evaluate a suite of augmentation strategies, finding that most hamper performance while time masking can reduce validation loss. A novel hybrid model, EEGNet-CharRNN, integrates a character-level RNN with EEGNet using a tunable weighting parameter to improve letter-level decoding, yielding up to 2.9 percentage-point gains on unseen subjects and about 96% accuracy on the all-subject evaluation at optimum weighting. These results demonstrate the potential for language-model priors to enhance BCI spelling in real-time, especially for subjects whose EEG signals are challenging to classify. The work outlines concrete directions for real-time full-sentence spellers and suggests that larger language models could further boost performance in practical deployments.

Abstract

Steady-State Visual Evoked Potential (SSVEP) spellers are a promising communication tool for individuals with disabilities. This Brain-Computer Interface utilizes scalp potential data from (electroencephalography) EEG electrodes on a subject's head to decode specific letters or arbitrary targets the subject is looking at on a screen. However, deep neural networks for SSVEP spellers often suffer from low accuracy and poor generalizability to unseen subjects, largely due to the high variability in EEG data. In this study, we propose a hybrid approach combining data augmentation and language modeling to enhance the performance of SSVEP spellers. Using the Benchmark dataset from Tsinghua University, we explore various data augmentation techniques, including frequency masking, time masking, and noise injection, to improve the robustness of deep learning models. Additionally, we integrate a language model (CharRNN) with EEGNet to incorporate linguistic context, significantly enhancing word-level decoding accuracy. Our results demonstrate accuracy improvements of up to 2.9 percent over the baseline, with time masking and language modeling showing the most promise. This work paves the way for more accurate and generalizable SSVEP speller systems, offering improved communication solutions for individuals with disabilities.

Improving SSVEP BCI Spellers With Data Augmentation and Language Models

TL;DR

This work tackles the variability and generalization challenges of SSVEP-based BCI spellers by coupling data augmentation with a language-model-based prior. The authors establish a strong EEGNet baseline on the Benchmark dataset and systematically evaluate a suite of augmentation strategies, finding that most hamper performance while time masking can reduce validation loss. A novel hybrid model, EEGNet-CharRNN, integrates a character-level RNN with EEGNet using a tunable weighting parameter to improve letter-level decoding, yielding up to 2.9 percentage-point gains on unseen subjects and about 96% accuracy on the all-subject evaluation at optimum weighting. These results demonstrate the potential for language-model priors to enhance BCI spelling in real-time, especially for subjects whose EEG signals are challenging to classify. The work outlines concrete directions for real-time full-sentence spellers and suggests that larger language models could further boost performance in practical deployments.

Abstract

Steady-State Visual Evoked Potential (SSVEP) spellers are a promising communication tool for individuals with disabilities. This Brain-Computer Interface utilizes scalp potential data from (electroencephalography) EEG electrodes on a subject's head to decode specific letters or arbitrary targets the subject is looking at on a screen. However, deep neural networks for SSVEP spellers often suffer from low accuracy and poor generalizability to unseen subjects, largely due to the high variability in EEG data. In this study, we propose a hybrid approach combining data augmentation and language modeling to enhance the performance of SSVEP spellers. Using the Benchmark dataset from Tsinghua University, we explore various data augmentation techniques, including frequency masking, time masking, and noise injection, to improve the robustness of deep learning models. Additionally, we integrate a language model (CharRNN) with EEGNet to incorporate linguistic context, significantly enhancing word-level decoding accuracy. Our results demonstrate accuracy improvements of up to 2.9 percent over the baseline, with time masking and language modeling showing the most promise. This work paves the way for more accurate and generalizable SSVEP speller systems, offering improved communication solutions for individuals with disabilities.
Paper Structure (20 sections, 3 equations, 7 figures, 2 tables)

This paper contains 20 sections, 3 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Baseline Model
  4. Data
  5. Benchmark Dataset
  6. Data Preprocessing
  7. Data Partitioning
  8. Baseline Implementation
  9. Model Extensions
  10. Data Augmentation
  11. Language Model
  12. Dataset Creation
  13. Hybrid Model Evaluation Metric
  14. Results and Discussion
  15. Data Augmentation
  16. ...and 5 more sections

Figures (7)

  • Figure 1: Overall visualization of the EEGNet architecturelawhern2018.
  • Figure 2: The average classification accuracies obtained by GDNet-EEG and five other baseline models including EEGNet wan2023.
  • Figure 3: CharRNN Diagram
  • Figure 4: Hybrid EEGNet-CharRNN Model
  • Figure 5: Table of ablations
  • ...and 2 more figures