Attention model of EEG signals based on reinforcement learning

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are computationally expensive for large Electroencephalogram (EEG) datasets due to linear complexity with signal dimensions.
• A new Gated Recurrent Unit (GRU) network model is proposed, integrating reinforcement learning (RL) for attention mechanisms in EEG signal processing.

Purpose of the Study:

• To develop a computationally efficient and accurate model for EEG signal processing.
• To adaptively extract relevant information from EEG signals at multiple scales and resolutions.
• To overcome the limitations of traditional CNNs in handling high-dimensional EEG data.

Main Methods:

• The proposed model employs RL to treat attention mechanism implementation as a learning problem.
• Policy gradient methods are utilized for end-to-end model training, addressing non-differentiability introduced by RL.
• The network exhibits translation invariance, making computational complexity independent of EEG signal dimensions.

Main Results:

• The model was evaluated on the BCI Competition IV-2a dataset.
• Achieved 86.78% accuracy in subject-dependent mode and 71.54% in subject-independent mode.
• Outperformed current state-of-the-art methods on the benchmark dataset.

Conclusions:

• Attention models combined with RL principles enhance EEG signal decoding accuracy.
• The proposed approach effectively filters noise and redundant data, focusing on key EEG features.
• This method offers a significant advancement in EEG signal processing for brain-computer interfaces.