Attention Temporal Convolutional Network With Weight Sharing for Motor Imagery Signal Classification
Year: 2025
Type: preprint
Abstract: Brain–computer interfaces (BCIs) offer a non-muscular communication pathway by enabling users to interact with external devices through brain signals alone. A widely adopted BCI paradigm is based on motor imagery (MI), where individuals imagine specific limb movements without actual muscle activity. Electroencephalography (EEG) is commonly employed to capture these MI-induced neural signals, which are then processed and classified to infer the user's intended action. However, the inherent low signal-to-noise ratio of EEG data, along with significant inter-subject variability, poses substantial challenges for accurate and robust classification. In this work, we propose the Attention Temporal Convolutional Network with Weight Sharing (ATCNet-WS) to enhance MI EEG classification performance. ATCNet-WS employs multi-head self-attention to capture discriminative spatio-temporal features and utilizes multi-layer dilated temporal convolutions to aggregate information across long temporal contexts. A sliding-window based multi-branch architecture is applied to provide additional network width. To mitigate overfitting, weight sharing is applied across temporal convolution layers with varying dilation rates. Cosine annealing learning rate schedule with warm restarts is adopted to accommodate the different training phases effectively. ATCNet-WS achieves subject-dependent four-class classification accuracy of 87.06% on the BCI Competition IV 2a dataset, and subject-dependent two-class classification accuracy of 87.85% on the BCI Competition IV 2b dataset, both outperforming state-of-the-art methods.
Authors: Zefeng Xu, Zhuliang Yu
Institution: South China University of Technology
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: gold
OpenAlex ID: W4413514493