Neuroscience Prior knowledge guided EEG representation disentanglement for auditory attention decoding.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Accurate Auditory Attention Decoding (AAD) requires disentangling attended speech from Electroencephalogram (EEG) signals.
• Existing AAD models often neglect crucial neuroscience priors like hierarchical processing and temporal asynchrony, limiting performance and interpretability.
• Unattended speech interference in EEG further complicates accurate AAD.

Purpose of the Study:

• To propose a novel neuroscience-inspired framework for AAD that explicitly incorporates critical brain processing priors.
• To enhance the disentanglement of attended speech components from EEG signals.
• To improve the interpretability and accuracy of Auditory Attention Decoding models.

Main Methods:

• Utilized EEGViT for hierarchical segmentation and integration of EEG signals, mirroring auditory information processing.
• Introduced Hierarchical Contrastive Learning (HCL) for fine-grained alignment between EEG and speech embeddings (using WavLM).
• Developed Hierarchical Mutual Information Minimization (HMIM) to disentangle attended from unattended speech components.

Main Results:

• The proposed framework significantly outperformed state-of-the-art AAD methods across three public datasets.
• Ablation studies validated the effectiveness of individual components (EEGViT, HCL, HMIM).
• Learned representations demonstrated consistency with neuroscience priors, confirming enhanced performance and interpretability.

Conclusions:

• The neuroscience-inspired framework offers a significant advancement in Auditory Attention Decoding.
• Explicitly incorporating brain processing priors improves AAD accuracy and model interpretability.
• The method provides a robust approach for disentangling complex auditory neural signals.