Related Experiment Video
Updated: Aug 30, 2025
06:34
A Single-Channel and Non-Invasive Wearable Brain-Computer Interface for Industry and Healthcare
Published on: July 7, 2023
2.5K
An efficient CNN-LSTM network with spectral normalization and label smoothing technologies for SSVEP frequency
Yudong Pan1, Jianbo Chen1, Yangsong Zhang1,2
1Laboratory for Brain Science and Medical Artificial Intelligence,School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang 621010, People's Republic of China.
Journal of Neural Engineering
|August 30, 2022
Summary
A novel deep learning network, SSVEPNET, improves accuracy for steady-state visual evoked potentials (SSVEPs) brain-computer interfaces (BCIs). This SSVEP DL NETwork enhances both intra- and inter-subject classification performance, offering a promising advancement.
Area of Science:
- Neuroscience and Biomedical Engineering
- Signal Processing and Machine Learning
Background:
- Steady-state visual evoked potentials (SSVEPs) offer high information transfer rates for brain-computer interfaces (BCIs).
- Current frequency recognition methods for SSVEPs struggle with performance dependency on calibration data, especially for inter-subject classification.
- Existing deep learning (DL) approaches for inter-subject classification show potential but require further performance enhancement.
Purpose of the Study:
- To propose an efficient deep learning network (SSVEPNET) for improved SSVEP frequency recognition.
- To enhance SSVEPNET performance using spectral normalization and label smoothing techniques.
- To evaluate SSVEPNET's accuracy for both intra- and inter-subject classification against baseline methods.
Main Methods:
- Developed SSVEPNET, a deep learning model integrating one-dimensional convolution and Long Short-Term Memory (LSTM) modules.
- Implemented spectral normalization and label smoothing to optimize the SSVEPNET architecture.
- Conducted comparative evaluations using two SSVEP datasets, varying time-window lengths (0.5s, 1s) and training data sizes.
Main Results:
- SSVEPNET consistently achieved the highest average accuracy for both intra- and inter-subject classification across all experimental conditions.
- The proposed model outperformed traditional and other deep learning baseline methods in SSVEP frequency recognition.
- Performance gains were observed across different datasets, time windows, and training data quantities.
Conclusions:
- The proposed SSVEPNET demonstrates significant promise for enhancing frequency recognition in SSVEP-based BCIs.
- The combination of convolutional neural network (CNN) and LSTM architectures is effective for electroencephalography (EEG) data.
- Spectral normalization and label smoothing serve as valuable optimization strategies for developing efficient BCI models.
