Domain transfer learning for multi-class EEG decoding based on lateral feature connected diffusion model

초록

Multiclass neural decoding holds significant potential for developing intuitive brain-computer interface systems that enable direct translation of neural activity into communication or control commands. However, the neural correlates of imagined speech are inherently weaker, making robust decoding, particularly with non-invasive electroencephalography (EEG), substantially more challenging. To address these issues, this study proposes a transfer learning framework based on knowledge distillation, designed to transfer rich linguistic representations extracted from overt speech EEG to the imagined speech domain. This approach aligns the representational gap between overt and imagined speech, improving the discriminability and stability of imagined speech features. We designed a diffusion-based deep neural network, incorporating a U-Net backbone with lateral feature connections to facilitate multi-scale feature fusion and effective representation learning. The proposed model demonstrates significant improvements in multi-class decoding performance and maintains stable, scalable performance across subjects. These findings indicate that integrating overt speech knowledge into imagined speech decoding constitutes an effective strategy for enhancing decoding reliability and generalization. The proposed framework provides a systematic pathway toward practical and accessible EEG-based communication systems. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드