Foundation and Large-Scale AI Models in Neuroscience: A Comprehensive Review

Shihao Yang, Xiying Huang, Danilo Bernardo, Jun-En Ding, Andrew Michael, Jingmei Yang, Patrick Kwan, Ashish Raj, Feng LiuarXiv 2025 · DOI · arXiv

Abstract

Neuroscience is witnessing a paradigm shift from task-specific, hand-engineered machine learning models to large-scale Foundation Models (FMs). This review explores the architectural transitions, data resources, and clinical implications of this new era, while addressing the critical "validation gap" that hinders real-world deployment.

Introduction

Neuroscience data presents unique challenges: it is high-dimensional, non-stationary, and spans multiple spatiotemporal scales—from millisecond electrical spikes to minute-long blood-oxygen changes in fMRI. Historically, researchers relied on manual feature engineering, a process limited by human intuition and poor generalizability. The emergence of Foundation Models—massive systems trained on unlabeled data—offers a path toward automatic, hierarchical representation learning of the brain's complex signals.

The Evolution of AI in Neuroscience

The trajectory of AI in brain research can be divided into three distinct eras. From 1980 to 2010, classical machine learning dominated, utilizing algorithms like SVMs, clustering, and regression on hand-crafted features. The subsequent decade (2010-2020) saw the rise of Deep Learning (CNNs, RNNs), which automated feature extraction but remained largely task-specific. Today, we have entered the era of Large-scale AI Models, characterized by universal representations and cross-modal capabilities.

Evolution of AI in Neuroscience from classical machine learning to large-scale foundation models.

Operational Classification and Methodology

To categorize the current landscape, we propose a three-tier classification framework:

  1. General-Purpose Foundation Models: Systems like BrainLM that aim for universal neural representations across modalities.
  2. Modality-Specific Base Models: Architectures optimized for specific signals, such as EEGFormer for electroencephalography.
  3. Task-Oriented Large-Scale Models: Models fine-tuned for specific clinical endpoints, such as seizure detection or disease classification.

Key architectural innovations driving this shift include Transformers for capturing long-range temporal dependencies, Self-Supervised Learning (SSL) for utilizing unlabeled data via "masked signal modeling," and Generative Diffusion Models for high-fidelity "brain reading" (reconstructing images or text from fMRI patterns).

Data Resources and Scaling Laws

The development of these models is fueled by massive open-access datasets. The TUH EEG Corpus provides 1.7 TB of data from over 15,000 subjects, while the UK Biobank offers MRI data from over 61,000 subjects. Research confirms that "Scaling Laws" apply to neuroscience; for instance, EEG model accuracy follows a power-law relationship ($ACC = (33.6 \times N)^{0.029}$) relative to parameter count, suggesting that performance continues to improve with model size and data volume.

Key Insights: The Validation Gap and Efficiency

A critical finding of this review is the "Validation Gap." Models that achieve state-of-the-art results on curated benchmarks often fail in clinical settings due to differences in hardware or patient populations. Addressing this requires "deployment-mimicking" evaluation strategies and a shift away from simple random train-test splits.

Conversely, significant progress has been made in data efficiency. The MindEye2 model can now achieve high-quality visual reconstruction using only one hour of subject-specific data—a 40x improvement over previous iterations that required dozens of hours of training per subject.

Results and Clinical Impact

Large-scale models are transforming core functions in neuroscience, from neural signal decoding to predictive modeling. Clinical LLMs like Almanac have demonstrated an 18-percentage-point improvement in factuality over general models in medical Q&A. In diagnostic applications, video-based transformers like SETR-PKD have achieved 83.9% accuracy in detecting tonic-clonic seizures non-invasively, offering a viable alternative to traditional EEG monitoring.

Core functions, cross-modal generative capabilities, and clinical/scientific impacts of large-scale AI in neuroscience.

Future Frontiers and Limitations

Despite these successes, challenges remain. "Negative transfer" can occur when pre-training on unrelated data (like natural images) conflicts with neural signal biases. Furthermore, as models grow in complexity, they often become "black boxes," sacrificing biological interpretability for performance. Future work must focus on prospective clinical trials, real-time closed-loop neurotechnology, and the development of "Digital Twin Brains" for personalized surgical and pharmacological simulation.

Summary updated

BibTeX
@misc{yang2025foundation,
  author = {Shihao Yang and Xiying Huang and Danilo Bernardo and Jun-En Ding and Andrew Michael and Jingmei Yang and Patrick Kwan and Ashish Raj and Feng Liu},
  title = {{Foundation and Large-Scale AI Models in Neuroscience: A Comprehensive Review}},
  year = {2025},
  eprint = {2510.16658},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.16658},
}
Figure