Research map / Architectures and efficiency / Mixture of experts
NeurIPT: Foundation Model for Neural Interfaces
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 NeurIPT, an EEG foundation model (Crossformer-based transformer with amplitude-aware masked pretraining, progressive mixture-of-experts, 3D electrode embedding, and intra-inter lobe pooling) pretrained on over 2,000 hours of public EEG and fine-tuned on eight BCI datasets.
Problem
Applying foundation models to EEG is hard because of inter-subject, inter-task, and inter-condition variability and heterogeneous electrode configurations; existing EEG FMs ignore 3D electrode geometry, rely on random masking that encourages local interpolation, and use fully connected or global poo
Approach
Mask EEG by signal amplitude instead of random contiguous intervals, add a progressive MoE whose expert count grows with depth, embed electrodes with sinusoidal encodings of their 3D coordinates, and pool features within and across brain lobes during fine-tuning.
Key results
- On MentalArithmetic, NeurIPT reaches 86.46 balanced accuracy versus 72.56 for the best baseline (CBraMod), a +13.90 gain.
- On TUAB, NeurIPT's balanced accuracy is 82.93, only +0.04 over CBraMod's 82.89, while its Cohen's kappa (90.40) and AUROC (89.49) are below CBraMod's 92.58 and 92.27.
- In a from-scratch ablation, amplitude-aware masking reaches 68.45% balanced accuracy on the SVM probe versus 63.20% for BERT-style random masking at 50% masking ratio.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use amplitude-aware instead of random masking during pretraining (ablation) | NeurIPT representations evaluated with SVM · TUEV | Balanced Accuracy: 0.6320 -> 0.6845; 67.35 -> 67.61 | Helped |
| Increase masking ratio from 30 to 50 (ablation) | NeurIPT representations evaluated with SVM · TUEV | Balanced Accuracy: 0.5523 -> 0.6845 | Helped |
| Increase masking ratio from 50 to 60 (ablation) | NeurIPT representations evaluated with SVM · TUEV | Balanced Accuracy: 0.6845 -> 0.5596 | Hurt |
| Use 3D electrode positional encoding instead of trigonometric encoding (ablation) | NeurIPT trained from scratch · TUEV | Balanced Accuracy: 67.72 -> 68.94 | Helped |
| Use IILP rather than no pooling (ablation) | NeurIPT trained from scratch · TUEV | Balanced Accuracy: 62.33 -> 68.94 | Helped |
| Use progressive rather than no experts (ablation) | NeurIPT trained from scratch · TUEV | Balanced Accuracy: 65.83 -> 68.94 | Helped |
| Use progressive rather than uniform expert allocation (ablation) | NeurIPT trained from scratch · TUEV | Balanced Accuracy: 65.91 -> 68.94 | Helped |
| Use progressive rather than shrinking expert allocation (ablation) | NeurIPT trained from scratch · TUEV | Balanced Accuracy: 65.80 -> 68.94 | Helped |
| Use progressive expert configuration [0,0,2,3,4,5] rather than [0,0,2,4,4,6] (ablation) | NeurIPT trained from scratch · BCIC-IV-2A | Balanced Accuracy: 44.01 -> 41.58 | Hurt |
| Use progressive expert configuration [0,0,2,4,6,8] rather than [0,0,2,4,4,6] (ablation) | NeurIPT trained from scratch · BCIC-IV-2A | Balanced Accuracy: 44.01 -> 41.15 | Hurt |
| Use progressive expert configuration [0,0,3,6,9,12] rather than [0,0,2,4,4,6] (ablation) | NeurIPT trained from scratch · BCIC-IV-2A | Balanced Accuracy: 44.01 -> 38.19 | Hurt |
| Use SwiGLU rather than ReLU (ablation) | NeurIPT trained from scratch · BCIC-IV-2A | Balanced Accuracy: 47.31 -> 44.01 | Hurt |
Surprising findings
- Random masking yields lower reconstruction loss than amplitude-aware masking, yet amplitude-aware masking gives better downstream classification, which the authors attribute to random masking encouraging interpolation.
- NeurIPT does not beat the best baseline on Cohen's kappa and AUROC for TUAB, and it has no gain over CBraMod in balanced accuracy beyond 0.04.
Limitations
- Some datasets show reinitialization or dataset-specific choices needed (e.g., SEED-V, BCIC-IV-2A), suggesting sensitivity to downstream adaptation.
- Ablations of MoE, pooling, and components are run on models trained from scratch rather than the pretrained model, and gains are small on several datasets (e.g., SEED-V +0.13).
Experimental setup
Datasets: MentalArithmetic Mumtaz2016 PhysioNetP300 Sleep-EDFx SEED-V BCIC-IV-2A TUAB TUEV
Benchmarks: MentalArithmetic Mumtaz2016 PhysioNetP300 Sleep-EDFx SEED-V BCIC-IV-2A TUAB TUEV
Metrics: balanced accuracy Cohen's kappa AUC-PR weighted F1 AUROC
Baselines: BENDR BIOT LaBraM EEGPT NeuroLM CBraMod EEGNet
Builds on: Crossformer CBraMod BERT Mixture-of-Experts FlashAttention-2
Closest papers by meaning
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.90)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.90)
- CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.89)
- Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.89)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.89)
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.