Related Experiment Video
Updated: Jun 8, 2025
06:34
A Single-Channel and Non-Invasive Wearable Brain-Computer Interface for Industry and Healthcare
Published on: July 7, 2023
2.3K
A hybrid local-global neural network for visual classification using raw EEG signals.
Shuning Xue1,2, Bu Jin2, Jie Jiang2
1The School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, 100049, China.
Scientific Reports
|November 7, 2024
Summary
This study introduces a novel hybrid local-global neural network for decoding visual information using electroencephalography (EEG) brain-computer interfaces (BCIs). The new model effectively processes raw EEG signals, outperforming existing methods on various datasets.
Area of Science:
- Neuroscience
- Machine Learning
- Biomedical Engineering
Background:
- Brain-computer interfaces (BCIs) show promise for decoding visual information from EEG signals.
- Artificial neural networks (ANNs) have been applied to classify EEG data, but often underperform traditional methods when extracting features from raw signals.
- Current ANN-based methods are typically evaluated on limited datasets at low sampling rates, restricting deep learning model potential.
Purpose of the Study:
- To develop an advanced neural network architecture for end-to-end training on raw EEG signals, eliminating the need for handcrafted features.
- To improve the performance of EEG-based BCIs in visual information decoding, particularly in high sampling rate settings.
- To create a robust deep learning framework capable of handling diverse EEG datasets and sampling rates.
Main Methods:
- A hybrid local-global neural network was designed, incorporating a reweight module for adaptive channel weighting.
- Local and global feature extraction modules were developed to capture essential EEG signal characteristics.
- Spatial integration and feature fusion modules were implemented to enhance information processing and feature extraction, especially at high sampling rates.
Main Results:
- The proposed model achieved state-of-the-art performance on established small-scale EEG datasets.
- It outperformed baseline methods on three under-explored large-scale datasets.
- Ablation studies confirmed the consistent performance improvement provided by the individual modules across different datasets and sampling rates.
Conclusions:
- The hybrid local-global neural network offers a robust end-to-end learning framework for EEG-based BCIs.
- The model effectively decodes visual information from raw EEG signals without manual feature engineering.
- The proposed architecture demonstrates significant improvements and adaptability across various datasets and sampling rates.
