Research map / Label-efficient and robust learning / Recognition and long-tail

What do EEG Foundation Models Capture from Human Brain Signals?

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Audits three EEG foundation models (CSBrain, CBraMod, LaBraM) on five clinical tasks against a 63-feature hand-crafted EEG lexicon, using ridge probing, LEACE-style subspace erasure, and a closure ratio with a transparent classifier.

Problem

Whether EEG foundation models, trained self-supervised on raw signals, encode and causally use the hand-crafted clinical EEG feature catalog, and how much of their task advantage that catalog explains.

Approach

Decomposes alignment into three questions (encoding via layer-wise ridge probes, usage via cross-covariance subspace erasure with null-target controls, and performance recovery via a closure ratio against a same-dimension random-feature baseline), applied to every (dataset, model, feature) triplet.

Key results

• 648 of 945 (model, task, feature) units (68.6%) are representation-causal and 199 (21.1%) encoded-only.
• Confirmed features recover on average 79.3% of the foundation-model advantage over a random-feature baseline (closure), from MDD ~0.99 down to Stress ~0.56.
• 50 features qualify as universal candidates with strong support in two or more tasks.

Surprising findings

• Frequency-domain features dominate usage, but the other five families each carry substantial causal mass on most tasks.
• Stress/CSBrain closure is 0.680 with only 6 representation-causal features.

Limitations

• Probe and eraser are linear, so causal reliance is restricted to the linearly decodable component.
• The eraser is a covariance-simplified approximation to LEACE rather than the full whitened operator.
• Lexicon is not exhaustive and some entries are GPU-friendly proxies for textbook estimators; findings limited to three models and five tasks.

Experimental setup

Datasets: Mumtaz MDD Mental Arithmetic stress ISRUC-Sleep TUSL seizure-type Siena seizure-detection

Benchmarks: Mumtaz MDD Mental Arithmetic stress ISRUC-Sleep TUSL Siena

Metrics: ROC-AUC macro-F1 R^2 probe strength closure ratio delta erase

Baselines: random-feature baseline minimal-spectrum baseline B0 all-63-feature classifier encoded-feature classifier family-supported classifier

Builds on: LEACE concept bottleneck models CBraMod LaBraM CSBrain

Closest papers by meaning

• NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.91)
• Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.91)
• CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.90)
• BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.89)
• CLEF: EEG Foundation Models for Learning Clinical Semantics (NeurIPS 2026 · similarity 0.89)
• NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.88)