Foundation Models for Neural Signal Decoding: EEG-Centered Perspectives Toward Unified Representations.
Jii Kwon1, Youmin Shin2,3
1Department of Brain & Cognitive Sciences, Seoul National University, Seoul, Republic of Korea.
The European Journal of Neuroscience
|December 29, 2025
Summary
Foundation models (FMs) offer robust neural decoding by learning from diverse brain data like EEG. This review highlights key design principles for developing effective and interpretable foundation models for neural signals.
Area of Science:
- Neuroscience and Artificial Intelligence
- Computational Neuroscience
- Machine Learning for Neural Data
Background:
- Neural recordings (EEG, ECoG, intracortical) provide insights into brain dynamics but face decoding challenges like high dimensionality and variability.
- Traditional machine learning and deep learning models struggle with generalizability and interpretability for complex neural data.
- Foundation models (FMs), pretrained on large datasets, show promise for creating robust, transferable, and physiologically grounded neural representations.
Purpose of the Study:
- To synthesize emerging foundation model (FM) approaches for neural decoding, focusing on Electroencephalography (EEG) as a primary platform.
- To critically examine representative EEG-based FM architectures and identify essential design principles.
- To discuss the extension of FM frameworks to other neural recording types like ECoG and intracortical signals.
Main Methods:
- Review and synthesis of current literature on foundation models applied to neural decoding.
- Analysis of EEG-based FM architectures, including Patched Brain Transformer, CBraMod, and BrainGPT.
- Identification of key design principles: physiology-aware representation learning, structure-aware architectures, and interpretability mechanisms.
Main Results:
- EEG is identified as the most practical platform for FM development due to data availability and clinical relevance.
- Essential design principles include capturing oscillatory/dynamic structure, incorporating spatial/anatomical priors, and ensuring neuroscientific validity.
- Current FM approaches often underutilize spatial information and may inherit non-neural objectives.
Conclusions:
- Foundation models hold significant potential for advancing neural decoding across various modalities (EEG, ECoG, intracortical).
- Future development requires biologically informed objectives, structure-aware architectures, interpretable representations, and standardized data ecosystems.
- Unified neural representations across scales can be achieved by extending FM frameworks.
