LGFormer: integrating local and global representations for EEG decoding.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Electroencephalography (EEG) decoding faces challenges due to temporal variability and low signal-to-noise ratio.
• Convolutional Neural Networks (CNNs) capture local features but have limited receptive fields.
• Transformers capture global dependencies but require extensive data and resources, limiting efficiency for limited EEG datasets.

Purpose of the Study:

• To propose LGFormer, a hybrid network for efficient EEG decoding by learning both local and global representations.
• To combine the strengths of CNNs and transformers for multiscale perception in EEG signals.
• To develop a lightweight and computationally efficient model for EEG decoding.

Main Methods:

• LGFormer utilizes a deep attention module for global information extraction and dynamic CNN focus adjustment.
• A local-enhanced transformer integrates CNNs and transformers for multiscale perception.
• The model was evaluated on four public EEG datasets for motor imagery, cognitive workload, and error-related negativity decoding.

Main Results:

• LGFormer achieved state-of-the-art performance within 200 training epochs across multiple EEG decoding tasks.
• A novel spatial and temporal attention visualization method demonstrated LGFormer's ability to capture discriminative features.
• The model showed enhanced interpretability and insights into its decision-making process.

Conclusions:

• LGFormer demonstrates superior performance and high training efficiency for EEG decoding.
• The hybrid approach effectively captures both local and global EEG signal characteristics.
• LGFormer presents a versatile and practical solution for various EEG decoding applications.