Research map / Probabilistic and scientific modeling / Molecules and proteins

SPERA: Spherical Prior EEG Foundation Model with Geometry- and Frequency-Aware Latent Prediction

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Neuro-symbolic probabilistic layers

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.

Proposes SPERA, an EEG foundation model pretrained with latent-space JEPA plus a Legendre-polynomial spherical electrode-geometry prior and a relational spectral regularizer, pretrained on ~80k hours from 29k subjects.

Problem

General-purpose EEG representation learning is hard due to heterogeneous subjects, devices, and electrode montages, and input-space reconstruction objectives may learn noise and artifacts.

Approach

Combine JEPA latent prediction with a learnable Legendre-polynomial pairwise geometry bias in attention (plus anchor-based variant for full attention), a hybrid factorized/full attention backbone, and a KL-based relational spectral regularizer that aligns latent similarity structure with log-power spectra.</222>

Key results

• SPERA-Base achieves an average balanced accuracy of 0.7502 under LPFT across nine tasks, exceeding REVE-Base (0.7150) by 0.035.
• SPERA-Large reaches 0.7665 average balanced accuracy under LPFT, with largest significant gains on TUEV (+0.078) and BCIC2020-3 (+0.050) for SPERA-Base.
• Under linear probing, SPERA-Base (194M) exceeds REVE-Large (408M) by +0.013 average balanced accuracy.

Surprising findings

• Learnable pairwise MLP geometry bias underperforms the structured Legendre parameterization, and REVE's spatial encoding performs worse than the Legendre bias in the ablation.
• On FACED, CodeBrain outperforms SPERA on all metrics, the only task where a recent model leads SPERA on every metric.

Limitations

• Scaling only explored up to 396M parameters; billion-scale not investigated.
• Evaluated only on fixed-duration classification tasks; continuous decoding, regression and long-horizon prediction not tested.
• Demographic metadata not consistently available, limiting fairness analysis; spherical prior assumes surface electrodes.

Experimental setup

Datasets: TUH TDBRAIN MESA Siena I-CARE HSP HMS NMT OpenNeuro datasets HBN_NC MOABB datasets THINGS2

Benchmarks: BCIC2020-3 BCIC-IV-2a PhysioNet-MI FACED MAT ISRUC Mumtaz TUAB TUEV

Metrics: balanced accuracy Cohen's kappa weighted F1 AUC-PR AUROC

Baselines: EEGNet EEGConformer SPaRCNet ContraWR CNN-Transformer FFCL ST-Transformer EEG2Rep EEGDM BIOT LaBraM CBraMod

Builds on: JEPA I-JEPA LaBraM CBraMod REVE

Closest papers by meaning

• CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.86)
• NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.86)
• BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.85)
• Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.85)
• CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.85)
• What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.84)