人工智能
计算机科学
计算机视觉
运动表象
脑电图
模式识别(心理学)
可视化
图像处理
图像(数学)
特征提取
作者
Heng Liang,Fengming Zhao,Peiliang Gong,Yucheng Liu,Ziyu Jia
标识
DOI:10.1109/tnnls.2026.3732213
摘要
Motor imagery is an important experimental paradigm in brain–computer interfaces (BCIs), with significant application potential and research value. Existing research primarily focuses on multichannel electroencephalography (EEG) signals, achieving strong performance. However, multichannel EEG models require high computational costs, making them challenging to deploy on wearable devices. In contrast, few-channel EEG models are more suitable for wearables, but their limited feature information hinders model performance. Knowledge distillation is an effective method for transferring knowledge. By utilizing the feature information from multichannel EEG models, we have effectively improved the performance of few-channel EEG models. Specifically, we design the multichannel multiview heterogeneous knowledge distillation (MMHKD) framework, which includes three modules: the multichannel knowledge distillation module, the mutual information (MI)-based knowledge distillation module, and the multiview knowledge distillation module. The multichannel knowledge distillation module transfers rich multichannel EEG feature information based on an adaptive attention mechanism. The MI-based knowledge distillation module maximizes the MI between heterogeneous models, alleviating the feature gap between them. The multiview knowledge distillation module transfers vital relational knowledge between views, further enhancing the effectiveness of knowledge distillation. Experimental results demonstrate that the MMHKD framework achieves state-of-the-art (SOTA) performance on BCI-IV-2a, OpenBMI, and SMR-BCI datasets, effectively improving the performance of few-channel EEG models.
科研通智能强力驱动
Strongly Powered by AbleSci AI