ZMW-RSVP: a time-frequency prior-guided normalization-free RSVP-BCI decoding model

Rapid Serial Visual Presentation (RSVP)-based brain-computer interfaces (BCIs) are valuable for target detection and rehabilitation because they require no motor involvement and can elicit reliable event-related potentials (ERP). However, single-trial EEG decoding remains challenging due to low signal-to-noise ratio and substantial inter-subject variability. To address the limitations of conventional approaches that rely on handcrafted features and to further explore internal feature transformation in single-trial ERP decoding, this paper proposes an RSVP-BCI decoding model, termed ZMW-RSVP, by extending a time-frequency Transformer framework with oscillatory gated attention and a normalization-free dynamic activation network. The proposed model incorporates neuroscience-inspired priors through a multi-band oscillatory gating mechanism, with an emphasis on theta-related activity associated with RSVP/P300 processing, and further guides attention allocation via a time-window bias. In addition, the model replaces internal LayerNorm modules with Dynamic Tanh, which provides an element-wise internal transformation without explicit mean-variance normalization and is evaluated as a task-related alternative for ERP-related discriminative feature modeling. Experiments on the Tsinghua RSVP dataset and the NeuBCI Target Retrieval RSVP-EEG dataset demonstrate that ZMW-RSVP achieves improved classification performance under the evaluated cross-subject settings, validating the effectiveness of the proposed approach for single-trial RSVP decoding.