Weighted multi-output randomized neural networks with a feature memory pool for electroencephalogram signal classification

Neural networks with a single output often suffer from limited robustness and consistency in decoding EEG signals across various subjects. In contrast, deep neural networks with multiple outputs—preserving hierarchical feature structures and promoting diversity in predictions—offer a promising alternative. To this end, this article proposes a novel multi-output randomized neural network (MORNN) for cross-subject motor imagery electroencephalogram (MI-EEG) classification. When an equal-contribution scheme is adopted, a few poor-performing output layers may dominate the final prediction and degrade overall performance. To address this issue, we introduce a dynamic weighting mechanism that suppresses inferior outputs by evaluating the compatibility of each sample with different output layers. In addition, a feature memory pool is employed to retain high-quality features from all hidden layers, thereby enriching the representational capacity of direct links. By incorporating these enhancements, we propose a weighted MORNN with feature memory (WMORNN-FM) for cross-subject EEG classification. Furthermore, a dual-projection domain adaptation method is applied to align feature distributions and mitigate inter-subject variability, thereby enhancing the decoding capability of WMORNN-FM. Experimental results demonstrate that WMORNN-FM surpasses state-of-the-art algorithms by 6%, providing a reliable solution for cross-subject MI-EEG decoding.