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

TL;DR

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.

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.

Figures (8)

• Figure 1: The overview of Uni-NTFM's core concepts and superior performance. The left panel summarizes the model's "Unified" principle, spatial topology strategy, and method of performance validation in a Q&A format. The two radar charts on the right demonstrate that Uni-NTFM comprehensively outperforms baseline models across nine distinct downstream tasks in both linear probing and fine-tuning settings.
• Figure 2: The end-to-end architecture of Uni-NTFM in detail. The data processing flow begins with the data augmentation of raw EEG signals, followed by Heterogeneous Feature Projection Module (HFPM) which parallelly decomposes the input EEG signals into three domain streams: time, frequency, and raw. Next, Dual-domain Cross-attention Module (DCM) performs cross fusion of time domain and frequency domain features, combined with Topological Embedding (TE) to encode the spatial prior information of the electrodes. Finally, the processed representations are sent to the core MoE-Trans. Block to learn universal semantic features.
• Figure 3: The visualization of ablation study on the TUAB and TUEV downstream tasks.
• Figure 4: This figure shows the scaling laws of the Uni-NTFM model, revealing a positive correlation between performance and both model size and data volume. The top row of plots indicates that with a fixed pre-training data size, model performance steadily improves as the number of parameters increases. The bottom row shows that for a fixed model size, performance also scales positively with more of pre-training data. All plots contrast the performance under "Pretrained Only" (blue curve) and "Pretrained - Fine-tuned" (red curve) evaluation settings.
• Figure 5: t-SNE Visualization of Learned Feature Representations. This figure provides a qualitative comparison of feature spaces learned on the BCIC-IV-2a (top row) and TUEV (bottom row) datasets. The columns represent features from different sources: (a, d) Raw Power Spectral Density (PSD) features, which serve as a baseline; (b, e) features extracted from a trained EEGNet model, representing a standard deep learning approach; and (c, f) features from our only pre-trained Uni-NTFM model. Each color corresponds to a distinct class within the respective dataset. The clear formation of well-separated and compact clusters in the rightmost column (c, f) visually demonstrates Uni-NTFM's superior ability to learn discriminative and generalizable neural representations.