A Few-Layer Multilayer Perceptron is Worth Attention for EEG Classification in Rapid Serial Visual Presentation Task.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Rapid serial visual presentation (RSVP) is key for electroencephalography (EEG)-based brain-computer interfaces (BCIs).
• Single-trial decoding in RSVP-EEG is challenging due to signal overlap and entanglement.
• Existing Transformer-based models like DisCo-Former show promise but suffer from attention collapse.

Purpose of the Study:

• To develop and evaluate DisCo-Former, a Transformer-based framework for RSVP-EEG decoding.
• To investigate the impact of architectural complexity versus signal structure modeling on BCI performance.
• To introduce DisCo-MLP, a simpler MLP-based variant, and compare its efficacy to DisCo-Former.

Main Methods:

• Developed DisCo-Former, integrating trend-periodicity disentanglement, channel-level embeddings, and contrastive learning.
• Introduced DisCo-MLP by removing the Transformer encoder from DisCo-Former.
• Evaluated both models on two datasets using three regimes, including within-subject decoding.

Main Results:

• DisCo-Former surpassed existing approaches but exhibited attention collapse.
• DisCo-MLP matched or outperformed DisCo-Former across datasets and evaluation regimes.
• Within-subject decoding achieved mean AUCs of 0.94-0.98, exceeding baselines.

Conclusions:

• Model effectiveness in RSVP-EEG decoding is driven by modeling signal structure rather than architectural complexity.
• Simpler models, guided by neurophysiological priors, offer a practical route to state-of-the-art EEG-based interfaces.
• DisCo-MLP demonstrates that complex architectures are not always necessary for high performance in RSVP-EEG decoding.