NeurIPT

NeurIPT (Neural Interfaces Pre-trained Transformer) is a foundation model for electroencephalography (EEG) aimed at brain-computer interfaces (BCIs). EEG-based neural decoding is notoriously difficult to generalize because signals vary across subjects, tasks, recording conditions, and—critically—across electrode montages that differ in number and placement. NeurIPT addresses this heterogeneity directly, learning both the shared (homogeneous) and the variable (heterogeneous) spatio-temporal structure of EEG so that a single pretrained backbone can be adapted to many downstream BCI paradigms.

The model was introduced by researchers from Xiamen University Malaysia, Columbia University, The Hong Kong Polytechnic University, Xiamen University, and Harbin Institute of Technology (Shenzhen), and was accepted to NeurIPS 2025. It sits among a growing family of EEG foundation models—alongside LaBraM, EEGPT, BIOT, and CBraMod—but distinguishes itself through pretraining and architectural choices tailored to the physical layout of electrodes and the amplitude characteristics of neural signals.

NeurIPT is pretrained self-supervised on a large corpus of unlabeled EEG and then fine-tuned on individual BCI tasks, following the now-standard pretrain-then-adapt recipe but with components designed specifically for the irregularities of real-world EEG acquisition.

#Key Features

• Amplitude-Aware Masked Pretraining (AAMP): Instead of masking random temporal intervals, AAMP masks tokens based on signal amplitude, encouraging the model to learn robust representations across varying signal intensities rather than relying on simple local interpolation.
• Progressive Mixture-of-Experts (PMoE): Specialized expert subnetworks are introduced progressively at deeper transformer layers (a [0, 0, 2, 4, 4, 6] configuration across six layers), letting the model adapt to the diverse temporal dynamics of EEG without a uniform, monolithic capacity.
• 3D electrode coordinate embedding: The model encodes the physical 3D position of each electrode, allowing it to generalize across montages with different channel counts and placements rather than assuming a fixed input layout.
• Intra-Inter Lobe Pooling: A spatial pooling scheme that aggregates information within and across brain lobes, capturing anatomically meaningful spatial structure in the scalp signal.

#Technical Details

NeurIPT is built on a six-layer transformer encoder augmented with the Progressive Mixture-of-Experts design described above. It was pretrained on more than 2,000 hours of EEG drawn from 14 public datasets—including Emobrain, the Grasp-and-Lift EEG Challenge, the Inria BCI Challenge, EEG Motor Movement/Imagery, the SEED series, and several Temple University Hospital corpora (TUAR, TUSZ, TUSL)—with the downstream evaluation datasets held out of pretraining. Evaluation spans eight BCI datasets covering distinct paradigms: MentalArithmetic (mental stress), Mumtaz2016 (mental disorder diagnosis), PhysioNet P300, Sleep-EDFx (sleep staging), SEED-V (emotion recognition), BCIC-IV-2A (motor imagery), TUAB (abnormal EEG detection), and TUEV (event classification). The model reports state-of-the-art results across these tasks, including 86.46% on MentalArithmetic (versus 72.56% for the prior best, CBraMod), 98.03% on Mumtaz2016, 55.04% on BCIC-IV-2A, 67.31% on PhysioNet P300, and 70.47% on Sleep-EDFx, generally outperforming baselines such as EEGPT and CBraMod.

#Applications

NeurIPT targets the full breadth of EEG-based neural interfaces. Because it ingests arbitrary electrode montages and adapts to new tasks via fine-tuning, it is useful for motor-imagery BCIs, P300 spellers, emotion and stress recognition, clinical sleep staging, and screening for abnormal or pathological EEG. Researchers building BCI systems benefit from a pretrained backbone that reduces the labeled-data burden of each new task, while clinical and affective-computing groups can reuse the same model across heterogeneous acquisition hardware without redesigning the input pipeline.

#Impact

By tying its pretraining objective to signal amplitude and its spatial modeling to real electrode geometry, NeurIPT advances how EEG foundation models handle the inter-subject, inter-task, and inter-montage variability that has long limited BCI generalization. Its consistent gains across eight diverse benchmarks position it as a strong reference point for subsequent EEG foundation-model work. A practical caveat at the time of writing: the public repository notes that the source code was accidentally deleted and is being recovered, and no pretrained weights have yet been released, so independent reproduction currently depends on that recovery.