Researchers proposed BELT, a modular Bayesian edge-cloud architecture that combines user-specific adaptation, lightweight classification and compressed data transfer. BELT-lite achieved 87.9% and 80.6% mean accuracy on the BCI Competition IV-2b and IV-2a datasets, respectively, and processed each sample in 6.75 milliseconds on an ARM Cortex-A7—21% faster than EEGNet but about 2.7 percentage points less accurate.
