Improving RSVP EEG Decoding Performance under Auditory Perturbation: A Dual-View Model with Robust Distillation
Objective:
Decoding event-related potentials (ERPs) from Rapid Serial Visual Presentation (RSVP) EEG offers a promising approach for target recognition in brain-computer interfaces (BCIs). However, auditory noise in real-world environments degrades target-ERP decoding, substantially reducing accuracy and robustness.
Methods:
We propose a novel time-frequency dual-view network (TFDV-Net) with temporal and time-frequency branches. Cross-view global features guide local feature extraction to fully exploit complementary information in both views and obtain a more complete representation of task-relevant ERP information from auditory-perturbed EEG. Additionally, dual-stream knowledge distillation transfers feature representations and inter-layer processing patterns from a clean-EEG-trained teacher through feature-distribution alignment and inter-layer relational constraints, guiding the student preserve task-relevant information under auditory perturbations.
Results:
Experiments used an RSVP-EEG dataset collected from nine participants under clean, urban-noise, and conversational-noise conditions. The proposed method, comprising the TFDV-Net backbone and the proposed dual-stream knowledge-distillation strategy, achieved balanced accuracies of 83.02% and 80.78% under the two noise conditions, respectively, approaching the 84.54% achieved by the TFDV-Net backbone under the clean condition. With the same dual-stream knowledge-distillation strategy, TFDV-Net outperformed the state-of-the-art baseline, TSformer, by 3.27% and 4.36% in BA under the two noise conditions, respectively.
Conclusion:
The proposed method combines dual-view feature extraction with dual-stream knowledge distillation, effectively improving RSVP-EEG decoding accuracy and robustness under auditory perturbations.
Significance:
This work provides a robust RSVP-EEG decoding approach for noisy environments, showing potential for application in nonlaboratory environments and offering insights into noise-robust decoding in other BCI paradigms.
