Semisupervised Deep Stacking Network with Adaptive Learning Rate Strategy for Motor Imagery EEG Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery electroencephalogram (EEG) applications face challenges with unlabeled data in supervised learning and lengthy pretraining.
• Manual feature extraction in EEG analysis is time-consuming and can lead to information loss.

Purpose of the Study:

• To propose a semisupervised deep stacking network with an adaptive learning rate strategy (SADSN) to address sample loss and manual feature extraction issues in EEG data.
• To accelerate the convergence of the contrastive divergence (CD) algorithm using an adaptive learning rate strategy within the SADSN framework.

Main Methods:

• Developed a semisupervised deep stacking network (SADSN) incorporating an adaptive learning rate strategy.
• Integrated prior knowledge into the deep stacking network's intermediary layer.
• Trained a restricted Boltzmann machine using a semisupervised method with performance-based adaptive learning rate adjustments.

Main Results:

• The SADSN demonstrated advanced recognition accuracy for motor imagery classification using EEG data.
• The proposed method achieved a more significant convergence rate compared to existing approaches.
• Successfully classified motor imagery tasks, indicating practical utility.

Conclusions:

• The SADSN effectively mitigates sample loss and reduces the need for manual feature engineering in EEG analysis.
• The adaptive learning rate strategy significantly accelerates model convergence.
• The SADSN offers a promising approach for practical motor imagery EEG applications.