SimNeXt-EEG: Compact Motor Imagery EEG Decoding via Parameter-Free Temporal Attention and Separable Refinement

Abstract

Deployment-oriented motor imagery (MI) decoding with non-invasive electroencephalography (EEG) requires compact brain-computer interface (BCI) models that preserve low-SNR, non-stationary, and session-/subject-variable sensorimotor rhythms without heavy spatio-temporal parameterization or learnable attention. To address this, we propose SimNeXt-EEG, an end-to-end compact decoder built from three signal-oriented operators: multi-scale temporal feature fusion (MSTF) for oscillatory temporal motor activity, spatio-temporal encoding (STE) for compact temporal feature integration across multiple brain regions, and spatio-temporal refinement (STR) for compact yet discriminative spatio-temporal representation. A key component is the one-dimensional Simple Attention Module (1D-SimAM), which adapts SimAM's parameter-free energy-based saliency estimation from image locations to EEG temporal traces, enabling trace-wise temporal recalibration without learnable attention parameters. SimNeXt-EEG was evaluated on three datasets: BCIC-IV-2a, BCIC-IV-2b, and OpenBMI. For BCIC-IV-2a and 2b, SimNeXt-EEG achieved 70.81% and 80.96% session-independent accuracy, respectively, improving upon recent high-capacity baselines while requiring only 1.84 K and 1.31 K parameters, with CPU latencies of 1.46 and 1.10 ms. These results demonstrate a practical accuracy-compactness trade-off for deployment-oriented MI-EEG decoding.