USEA

Public EEG corpora are almost never uniformly annotated. One archive carries an expert label on every segment, the next has labels for a handful of subjects, and a third is raw recording with no annotation at all. The prevailing recipe for EEG foundation models — unsupervised pretraining followed by supervised fine-tuning — handles that unevenness by splitting it across stages, and the authors' objection is that the split over-adapts to individual datasets: each fine-tune pulls the shared encoder toward one corpus's idiosyncrasies, and the classification head it produces transfers to nothing.

USEA removes the split. A single backbone is optimized in one pass over all nine datasets, each contributing whichever objective its labels can support. Where labels exist, the label description is written out in natural language, encoded by a frozen language model, and used as the semantic target: the EEG representation is pulled toward that embedding by cosine similarity. Where labels are absent, the same backbone is trained self-supervised, by autoregressively reconstructing EEG token representations. Both losses act on one set of weights in one optimization, so fully labeled, few-labeled, and unlabeled corpora sit in the same training run. Because classification is a similarity comparison rather than a learned output layer, there is no task-specific head to refit, and the candidate label set becomes an input supplied at inference.

The model was developed at the Beijing Institute of Technology, with a collaborator at Nanyang Technological University, and published in Neural Networks. It joins a line of EEG-language models that includes NeuroLM, which quantizes EEG into discrete text-aligned tokens for a language model, and BLPM, which also matches EEG against label-text embeddings but keeps the two-stage structure USEA argues against. The acronym is shared with the Utah Slanted Electrode Array, an implanted peripheral-nerve electrode array unrelated to this work.

#Key Features

• One backbone, one optimization: Labeled, partially labeled, and unlabeled datasets are absorbed in a single training framework rather than in separate pretraining and fine-tuning passes, so no dataset gets its own copy of the encoder.
• Label text as the supervision signal: Class labels are expressed as natural-language descriptions and encoded by a frozen language model, so semantically related labels from different datasets reinforce each other instead of being treated as unrelated classes.
• No task-specific classification heads: Decoding is a cosine-similarity lookup against candidate label embeddings, so a new label vocabulary is provided at inference rather than trained in.
• Autoregressive reconstruction for unlabeled data: Unannotated recordings train the same backbone by predicting EEG token representations, keeping the objective consistent across corpora with different annotation budgets.
• Normalization-free transformer: The backbone replaces normalization layers with Dynamic Tanh, reducing sensitivity to dataset-specific activation statistics while preserving the amplitude information that carries meaning in EEG.

#Technical Details

USEA formulates EEG modeling as semantically guided autoregressive representation learning. The backbone is a transformer built on Dynamic Tanh (DyT), an elementwise squashing function with a learnable scale that substitutes for the normalization layers of a standard transformer block. The motivation is specific to EEG: normalization statistics encode the amplifier gain, montage, and referencing of whichever corpus a batch came from, and normalizing them away discards signal amplitude, which is diagnostic in EEG rather than nuisance. The supervised branch encodes each label description with a frozen language model and aligns EEG representations to it by cosine similarity; the unsupervised branch reconstructs EEG token representations autoregressively. Evaluation spans nine public EEG datasets covering different experimental paradigms and recording configurations, on which the authors report performance competitive with state-of-the-art baselines rather than a claim of leading them.

#Applications

The design targets groups whose EEG holdings are heterogeneous by default — a clinical archive with partial annotation, a few small labeled task recordings, and a larger pool of unlabeled sessions. Rather than deciding which subset justifies a pretraining run and which deserves a fine-tune, a lab can put all of it into one optimization and let annotation status determine the loss. The head-free decoding layer matters most when the label set is not fixed in advance: adding a class, renaming one, or evaluating on a cohort labeled under a different clinical vocabulary is a change to the text supplied at inference, not a retraining job.

#Impact

USEA's contribution is structural rather than a leaderboard result. It argues that the two-stage pretrain-then-fine-tune convention, inherited from vision and language, is a poor fit for a modality whose datasets differ in label availability as much as in electrode layout, and shows that one unified semi-supervised objective can cover the range. Pairing that with label-semantic supervision — which LaBraM and CBraMod do without — lets supervision from one dataset inform tasks it was never labeled for. Evaluation is in-silico, on retrospective public datasets, with no prospective or clinical validation reported, and no code, trained weights, or model card have been released.