Uni-NTFM: A Unified Foundation Model for EEG Signal Representation Learning

Zhisheng Chen, Yingwei Zhang, Qizhen Lan, Tianyu Liu, Huacan Wang, Yi Ding, Ziyu Jia, Ronghao Chen, Kun Wang, Xinliang Zhou

Abstract

Current foundation models for electroencephalography (EEG) rely on architectures adapted from computer vision or natural language processing, typically treating neural signals as pixel grids or token sequences. This approach overlooks that the neural activity is activated by diverse sparse coding across a complex geometric topological cortex. Inspired by biological neural mechanisms, we propose the Unified Neural Topological Foundation Model (Uni-NTFM), an architecture rooted in three core neuroscience principles. In detail, to align with the brain's decoupled coding mechanism, we design the Heterogeneous Feature Projection Module. This module simultaneously encodes both time-domain non-stationary transients and frequency-domain steady-state rhythms, ensuring high quality in both waveform morphology and spectral rhythms. Moreover, we introduce a Topological Embedding mechanism to inject structured spatial priors and align different sensor configurations onto a unified latent functional topography, effectively reconstructing the geometry of brain regions. Furthermore, we achieve functional modularization and sparse coding efficiency of biological networks by constructing the Mixture-of-Experts Transformer network. This dynamic routing mechanism assigns different signal patterns and tasks to specialized neural subnetworks, and effectively preventing task interference while increasing the model capacity to record-breaking 1.9 billion parameters. Uni-NTFM is pre-trained on a diverse corpus comprising 28,000 hours of EEG data, and outperforms existing models across nine distinct downstream tasks under both linear probing and fine-tuning settings, demonstrating that aligning model architecture with neural mechanisms is significant to learn universal representations and achieve generalizable brain decoding.

Introduction

Despite the success of foundation models in vision and language, applying this paradigm to Electroencephalography (EEG) has been challenging. Most current models treat EEG signals as simple pixel grids or token sequences, ignoring the unique physiological properties of the brain. The authors identify three critical "mismatches" in existing architectures: conflated coding (failing to separate time-domain transients from frequency rhythms), spatial neglect (ignoring the 3D geometric layout of the cortex), and computational inefficiency when scaling to varied neural patterns.

Methodology

The Unified Neural Topological Foundation Model (Uni-NTFM) addresses these gaps through three core biologically-grounded modules designed to mirror neural processing:

  • Heterogeneous Feature Projection Module (HFPM): Instead of uniform tokenization, it parallelly encodes signals into three streams: a Time Path for non-stationary transients, a Frequency Path for rhythmic patterns, and a Raw Path as a high-fidelity reference. This aligns with the brain's decoupled coding mechanisms.
  • Topological Embedding (TE): This hierarchical mechanism injects spatial priors (region-level, intra-region, and absolute position) into the model. It effectively maps diverse electrode layouts onto a unified "latent functional topography," allowing the model to understand the brain's geometry regardless of the specific sensor setup used.
  • Mixture-of-Experts (MoE) Transformer: To scale efficiently, standard dense layers are replaced with a sparsely-activated MoE block. This allows the model to reach a record-breaking 1.9 billion parameters while only activating approximately 307 million parameters per signal during inference, simulating the functional specialization of biological neural networks.
The end-to-end architecture of Uni-NTFM, featuring multi-domain feature projection and MoE blocks.

Experimental Setup

Uni-NTFM was pre-trained on a massive aggregated corpus of 28,000 hours of EEG data from over 17,000 subjects, including clinical recordings (TUEG, CAUEEG), emotion tasks (SEED), and resting-state data. For evaluation, the model was tested across nine distinct downstream tasks covering clinical diagnostics (abnormal detection, event classification), cognitive/emotion assessment, and Brain-Computer Interface (BCI) applications like motor imagery and sleep staging.

Results and Discussion

Uni-NTFM sets a new state-of-the-art (SOTA) across all nine downstream tasks in both linear probing and fine-tuning settings.

Performance and Scaling

On the 3-class SEED emotion task, Uni-NTFM-large achieved 73.37% balanced accuracy, outperforming previous foundation models like LaBraM-Base (73.18%) and traditional task-specific models. The authors observed that performance scales positively with both parameter count and data volume, though they noted that billion-parameter models require massive datasets to avoid saturation.

Radar charts comparing Uni-NTFM against baselines in multi-task (pretrained) and single-task (fine-tuned) settings across 9 tasks.

Robustness and Efficiency

The Topological Embedding allows the model to generalize across different electrode standards (e.g., clinical 10-20 vs. high-density 10-10 systems) without task-specific re-training. In "stress tests" involving random electrode dropouts, Uni-NTFM demonstrated significantly slower performance decay than models lacking spatial priors, proving its ability to infer missing information from neighboring brain regions. Furthermore, the MoE architecture ensures that despite its 1.9B parameter size, inference costs remain manageable and competitive with smaller dense models.

Conclusion

Uni-NTFM demonstrates that a "neuroscience-first" design—aligning model architecture with biological mechanisms like decoupled coding and cortical geometry—is key to learning universal neural representations. Future work will explore multi-modal integration, combining EEG with fMRI or eye-tracking to create more generalizable brain-computer interfaces.

Summary updated

Abstract

Current foundation models for electroencephalography (EEG) rely on architectures adapted from computer vision or natural language processing, typically treating neural signals as pixel grids or token sequences. This approach overlooks that the neural activity is activated by diverse sparse coding across a complex geometric topological cortex. Inspired by biological neural mechanisms, we propose the Unified Neural Topological Foundation Model (Uni-NTFM), an architecture rooted in three core neuroscience principles. In detail, to align with the brain's decoupled coding mechanism, we design the Heterogeneous Feature Projection Module. This module simultaneously encodes both time-domain non-stationary transients and frequency-domain steady-state rhythms, ensuring high quality in both waveform morphology and spectral rhythms. Moreover, we introduce a Topological Embedding mechanism to inject structured spatial priors and align different sensor configurations onto a unified latent functional topography, effectively reconstructing the geometry of brain regions. Furthermore, we achieve functional modularization and sparse coding efficiency of biological networks by constructing the Mixture-of-Experts Transformer network. This dynamic routing mechanism assigns different signal patterns and tasks to specialized neural subnetworks, and effectively preventing task interference while increasing the model capacity to record-breaking 1.9 billion parameters. Uni-NTFM is pre-trained on a diverse corpus comprising 28,000 hours of EEG data, and outperforms existing models across nine distinct downstream tasks under both linear probing and fine-tuning settings, demonstrating that aligning model architecture with neural mechanisms is significant to learn universal representations and achieve generalizable brain decoding. Our code is available at: https://anonymous.4open.science/r/Uni-NTFM-0924.
BibTeX
@inproceedings{chen2026uni,
  author = {Zhisheng Chen and Yingwei Zhang and Qizhen Lan and Tianyu Liu and Huacan Wang and Yi Ding and Ziyu Jia and Ronghao Chen and Kun Wang and Xinliang Zhou},
  title = {{Uni-NTFM: A Unified Foundation Model for EEG Signal Representation Learning}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=oUMiuYHW21},
}
Figure