STEAM: A Spatio-TEmporal Alignment Mixture-of-Experts Model with Hierarchical Pre-training for EEG Decoding

EEG decoding is hindered by inter-subject variability, low signal-to-noise ratios, and the high cost of training models from scratch for every new task or device. While foundation models have emerged to address these challenges, they often struggle to balance broad, cross-paradigm transferability with the high accuracy required for specific neural engineering applications. This paper asks: can a hierarchical framework effectively bridge the gap between general-purpose representation learning and paradigm-specific specialization?

The authors introduce STEAM (Spatio-TEmporal Alignment Mixture-of-Experts), a framework that employs a dual-branch encoder to process spatial (channel-wise) and temporal (segment-wise) information separately. These branches are coupled via a shared soft mixture-of-experts (SSMoE) module, which uses learnable soft slots to exchange information between branches without collapsing their distinct structural roles. The model also incorporates a frequency-aware gated attention mechanism to inject spectral priors directly into the attention layers.

STEAM is trained in two stages:

1. Stage-I: Large-scale self-supervised pre-training using masked reconstruction and cross-view contrastive learning to build a general backbone.
2. Stage-II: Continual pre-training on specific paradigms (e.g., motor imagery) using supervised objectives to refine decision boundaries.

Downstream adaptation is performed using a lightweight strategy that freezes the majority of the encoder and only updates the embedding modules and the task-specific head, significantly reducing computational overhead.

STEAM achieves the best average rank across seven diverse downstream datasets and fourteen evaluation settings, outperforming existing EEG foundation models while maintaining competitive inference costs. The hierarchical pre-training strategy (Stage-II) consistently improves accuracy over the general-purpose Stage-I initialization. Furthermore, the lightweight adaptation strategy proves highly effective, achieving performance comparable to full fine-tuning while updating only about 5% of the model's parameters.