STEAM

Electroencephalography is the workhorse of non-invasive brain-computer interfaces (BCIs), but decoding it is hard: signal-to-noise ratios are low, responses vary sharply between subjects, and montages, sampling rates, and paradigms differ from dataset to dataset. EEG foundation models promise a reusable backbone, yet face a three-way tension. Broad cross-paradigm pretraining often fails to beat small specialist networks on any single task; specializing to one paradigm usually means a new backbone from scratch; and adapting a pretrained encoder typically requires full fine-tuning, brittle when labeled trials are scarce.

STEAM (Spatio-TEmporal Alignment Mixture-of-Experts) is a hierarchical transfer framework built to reconcile those three goals. Its encoder splits into two branches — channel-wise tokens modeling spatial dependencies across electrodes, and temporal-segment tokens modeling dynamics within a trial — kept independent through the lower layers, then coupled at selected upper layers by a shared soft mixture-of-experts (SSMoE) module. SSMoE routes tokens from both branches into a small pool of soft expert slots and redistributes the fused result, letting the two views exchange information through a compact bottleneck without collapsing into a single stream. A frequency-aware gate additionally injects band-power priors into every encoder layer.

The model comes from Dongrui Wu's group at Huazhong University of Science and Technology, with the Beijing Zhongguancun Academy, and was posted as a preprint in August 2026. It is benchmarked head-to-head against EEG foundation models including LaBraM, CBraMod, Neuro-GPT, LUNA, and structurally informed encoders in the vein of CSBrain.

#Key Features

• Dual-branch spatio-temporal encoder: Separate spatial and temporal transformer stacks preserve channel-centric and segment-centric structure, with interaction deliberately delayed to the upper layers.
• Shared soft mixture-of-experts: Eight expert slots aggregate and redistribute tokens from both branches, so every token interacts with multiple slots — better suited to distributed EEG patterns than top-k routing.
• Frequency-aware gated attention: Per-token log band-power descriptors gate the attention output at each layer, embedding spectral evidence throughout the encoder rather than only at the head.
• General-to-paradigm hierarchical pretraining: A Stage-II continual-pretraining pass specializes the shared Stage-I backbone toward motor imagery or emotion recognition without retraining a separate model.
• Parameter-efficient adaptation: Freezing the backbone and updating only the 0.52-million-parameter embedding modules — about 5.1% of the model — plus a fresh task head stays competitive with full fine-tuning on classification.

#Technical Details

The Stage-I model holds 10.23 million parameters, reducing to a 9.43-million-parameter inference encoder once mask tokens and reconstruction decoders are dropped. Both branches use a hidden dimension of 128 with 8 attention heads; the SSMoE stack pairs 8 experts one-to-one with 8 slots and takes 61.9% of the parameter budget, though inference overhead stays small because communication passes through only eight slots. Inputs are resampled to 250 Hz and mapped to a 60-channel montage. Stage-I pretraining runs on the Temple University Hospital EEG Corpus — 6,451 recordings yielding 1,640,700 four-second segments — combining masked reconstruction at a 50% masking ratio, a token-diversity regularizer, and a symmetric InfoNCE loss aligning the two views of a trial, on a single NVIDIA A100. Stage-II adds supervised classification over a 45,195-trial motor-imagery corpus from seven datasets, or a 101,568-segment emotion corpus from SEED-V and DEAP.

Evaluation covers seven held-out datasets and fourteen settings, none appearing in either pretraining stage. Under cross-subject leave-one-subject-out testing, STEAM is best or tied-best on BNCI2014009, CHB-MIT, SEED, EEGMAT, and SEED-VIG, including 77.90% balanced accuracy on EEGMAT against 72.62% for the strongest baseline. In the within-subject few-shot protocol it reaches 92.90% on EEGMAT and 81.79% on BNCI2015001, and the emotion-specialized Stage-II variant lifts SEED from 61.21% to 65.18%. Lightweight adaptation beats full fine-tuning on three of six classification datasets and trails by at most 1.25 points on the others, but degrades SEED-VIG regression RMSE from 0.1625 to 0.2037.

#Applications

STEAM targets the practical bottleneck in BCI deployment: getting a usable decoder for a new subject, headset, or task without a long labeled calibration session. The frozen-backbone path lets one pretrained checkpoint serve motor-imagery control, P300 spellers, seizure detection, workload monitoring, emotion recognition, and vigilance estimation, updating only a small embedding stack and task head per deployment. Stage-II suits groups working within a single paradigm who want a specialized encoder without pretraining their own backbone.

#Impact

STEAM's contribution is less a new pretraining objective than a demonstration that generality and specialization need not be separate models: one 10-million-parameter backbone is continued into a paradigm-specific variant, then adapted with a fraction of its weights. That argument matters in a subfield where several published EEG foundation models underperform compact convolutional specialists, a pattern visible in the paper's own tables. The work is a preprint awaiting peer review, and no code or weights have been released, so the reported ranks cannot yet be independently reproduced. Results come from offline benchmark data rather than online BCI use, and the parameter-efficient adaptation carrying the efficiency claim is measurably weaker on regression than on classification.