BCIwiki (bciwiki.com) — Researchers from Tianjin University, the Haihe Laboratory of Brain-Computer Interaction and Human Machine Integration and the Southern University of Science and Technology have reviewed strategies for shrinking motor imagery brain-computer interface (BCI) decoders, an approach they term 'lightweight motor imagery (LightMI),' in a paper reported in Journal of Intelligent Medicine on September 26, 2026.
Motor imagery BCIs decode movement intentions from EEG signals to let the brain communicate directly with devices. Deep learning has substantially improved decoding accuracy, the authors note, but high computational complexity and large parameter counts make the models hard to deploy on resource-constrained hardware.
The review covers 2 areas. On the structural side, it examines efficient convolutional architectures, including depthwise separable, multi-branch and dilated convolutions, along with attention mechanisms. On model compression, it covers pruning, quantization, knowledge distillation and low-rank decomposition. Representative lightweight models are compared by parameter count, computational cost and classification accuracy.
The authors conclude that lightweight convolutional neural networks show strong potential for motor imagery BCIs, but each method has drawbacks: pruning can discard critical spatiotemporal features, low-bit quantization degrades performance, knowledge distillation depends heavily on teacher models, and low-rank decomposition adapts poorly across architectures. Manual design is also inefficient and hardware-specific. They call for work on automated neural architecture search, differentiable search and hardware-software co-design to make wearable motor imagery BCI systems practical.