Deep Unsupervised Representation Learning for Feature-Informed EEG Domain Extraction.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning models for electroencephalography (EEG) classification struggle with limited target subject data, hindering robust training.
• Existing transfer learning methods often assume a single distribution for latent features, failing to account for significant inter- and intra-subject signal variations.
• This leads to unstable decoding performance and poor model generalization due to unaddressed domain differences in EEG data.

Purpose of the Study:

• To propose a novel inference model, the Joint Embedding Variational Autoencoder (JE-VAE), for more accurate EEG classification.
• To address the challenge of inter- and intra-subject variability in EEG signals by improving latent feature distribution approximation.
• To enhance the optimization and scalability of deep learning models in EEG classification without sacrificing model tightness.

Main Methods:

• Developed a Joint Embedding Variational Autoencoder (JE-VAE) model utilizing jointly optimized variational autoencoders.
• Incorporated data-dependent inputs as an additional variable for improved model optimization.
• Demonstrated that maximizing the marginal log-likelihood of the second embedding section is key to learning the variational bound and achieving tighter lower bounds.

Main Results:

• The proposed JE-VAE model achieves conditionally tighter approximation of spatiotemporal feature distributions.
• The model demonstrates state-of-the-art performance in EEG data reconstruction and deep feature extraction.
• Analysis of extracted EEG signal domains provides insights into the reasons for disparities in subject adaptation efficacy.

Conclusions:

• The JE-VAE offers a robust solution for EEG classification with limited data by effectively managing subject-specific variations.
• The model's ability to approximate feature distributions more tightly leads to improved stability and generalization.
• This approach advances deep learning applications in neuroscience by providing better tools for analyzing complex EEG data.