Contrastive representation learning with transformers for robust auditory EEG decoding.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Decoding continuous speech from electroencephalography (EEG) is crucial for understanding auditory processing and developing hearing diagnostics.
• Deep learning has advanced EEG decoding, but low signal-to-noise ratios remain a challenge.

Purpose of the Study:

• To explore contrastive learning for robust latent representations of EEG signals.
• To introduce a novel model combining contrastive learning and transformer networks for auditory EEG decoding.
• To evaluate the model on auditory stimulus classification and envelope decoding tasks.

Main Methods:

• Utilized contrastive learning, a self-supervised technique, to learn EEG signal representations.
• Developed a novel architecture integrating contrastive learning with transformer networks.
• Evaluated the model on the ICASSP 2023 Auditory EEG Decoding Challenge tasks: match-mismatch classification and stimulus envelope decoding.

Main Results:

• Achieved state-of-the-art performance on both auditory EEG decoding tasks.
• Reached 87% accuracy in match-mismatch classification and a 0.176 Pearson correlation in envelope regression.
• Outperformed previous leading models in the challenge.

Conclusions:

• Contrastive learning shows significant potential for advancing auditory EEG decoding.
• The findings highlight the effectiveness of the proposed model for capturing relationships between auditory stimuli and EEG responses.
• The study provides insights into factors influencing model generalizability and accuracy for clinical applications.