Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
194 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Brain-JEPA: Brain Dynamics Foundation Model with Gradient Positioning and Spatiotemporal Masking (2409.19407v1)

Published 28 Sep 2024 in q-bio.NC, cs.AI, and cs.CV

Abstract: We introduce Brain-JEPA, a brain dynamics foundation model with the Joint-Embedding Predictive Architecture (JEPA). This pioneering model achieves state-of-the-art performance in demographic prediction, disease diagnosis/prognosis, and trait prediction through fine-tuning. Furthermore, it excels in off-the-shelf evaluations (e.g., linear probing) and demonstrates superior generalizability across different ethnic groups, surpassing the previous large model for brain activity significantly. Brain-JEPA incorporates two innovative techniques: Brain Gradient Positioning and Spatiotemporal Masking. Brain Gradient Positioning introduces a functional coordinate system for brain functional parcellation, enhancing the positional encoding of different Regions of Interest (ROIs). Spatiotemporal Masking, tailored to the unique characteristics of fMRI data, addresses the challenge of heterogeneous time-series patches. These methodologies enhance model performance and advance our understanding of the neural circuits underlying cognition. Overall, Brain-JEPA is paving the way to address pivotal questions of building brain functional coordinate system and masking brain activity at the AI-neuroscience interface, and setting a potentially new paradigm in brain activity analysis through downstream adaptation.

Summary

  • The paper introduces Brain Gradient Positioning to improve signal-to-noise ratio and capture subtle fMRI patterns using diffusion maps.
  • The paper presents Spatiotemporal Masking as a novel training strategy that dynamically adjusts observation ratios for superior model performance.
  • The paper demonstrates state-of-the-art accuracy in demographic prediction, disease diagnosis, and trait modeling across diverse datasets.

Brain-JEPA: A Comprehensive Review

The paper "Brain-JEPA: Brain Dynamics Foundation Model with Gradient Positioning and Spatiotemporal Masking" presents an innovative methodology for the analysis of brain activity data, with a focus on functional magnetic resonance imaging (fMRI). The authors introduced "Brain-JEPA," a foundation model that leverages the Joint-Embedding Predictive Architecture (JEPA) to predict abstract representations from brain activity observations. This paper demonstrates state-of-the-art performance in tasks such as demographic prediction, disease diagnosis/prognosis, and trait prediction, surpassing existing models like BrainLM.

Key Contributions

  1. Brain Gradient Positioning:
    • Description: The paper introduces Brain Gradient Positioning as an innovative method for determining the functional coordinates of brain regions. It leverages functional connectivity gradients to encode positional embeddings. Unlike traditional sine and cosine embeddings or anatomical positions, Brain Gradient Positioning captures brain functional parcellation more accurately, enhancing the model's ability to understand heterogeneous fMRI data.
    • Implementation: Gradients are derived using a non-negative affinity matrix and diffusion maps to create a continuous measure reflecting the functional relationships among different Regions of Interest (ROIs).
    • Outcome: This method provides a higher signal-to-noise ratio (SNR) and effectively captures subtle patterns, improving the robustness and scalability of the model.
  2. Spatiotemporal Masking:
    • Description: Spatiotemporal Masking, a tailored masking strategy, addresses the unique spatiotemporal characteristics of fMRI data. Instead of random block selection, the data is divided into Cross-ROI, Cross-Time, and Double-Cross regions. This ensures a diverse and challenging training regimen.
    • Implementation: The model predicts representations of multiple target blocks from a single observation block. Overlapped sampling is employed to adjust the observation-to-input ratio dynamically, fostering a stronger inductive bias.
    • Outcome: This approach significantly improves training efficiency and task performance, particularly in predicting brain activity representations.

Experimental Results

Performance Evaluation

Brain-JEPA excels across various tasks and datasets. Tables 1-3 showcase its superior performance in internal tasks (age and sex prediction) using a held-out portion of the UK Biobank dataset and external tasks involving other datasets such as HCP-Aging, ADNI, and MACC.

  • Demographic Prediction: Brain-JEPA outperformed previous models in age and sex predictions, achieving higher accuracy and F1 scores.
  • Disease Diagnosis/Prognosis: In tasks such as NC/MCI classification and amyloid positivity/negativity prediction, Brain-JEPA achieved state-of-the-art results, demonstrating strong potential for clinical applications.
  • Trait Prediction: Brain-JEPA was effective in predicting psychological traits like Neuroticism and cognitive functions like the Flanker score.

Performance Scaling

The performance scaling results, detailed in Figure 2, show that larger models like ViT-L consistently outperform smaller models such as ViT-S and ViT-B. This indicates the model's scalability with increasing size and data, highlighting its robustness.

Impact and Implications

Theoretical Implications

Brain-JEPA fosters a new paradigm in self-supervised learning for brain activity modeling. By employing the JEPA framework, the model provides more generalized and highly abstract representations of fMRI data, crucial for advancing the understanding of neural circuits underlying cognition.

Practical Implications

The model's application extends to diverse real-world scenarios, from clinical settings for disease prognosis to neuroscience research for understanding cognitive processes. Brain-JEPA's success across different ethnic groups underscores its potential in global healthcare applications, enhancing the inclusivity and applicability of AI-driven medical technologies.

Future Directions

Several avenues merit further exploration:

  • Larger Models: Pursuing larger models like ViT-H could yield even better performance.
  • Diverse Datasets: Incorporating a broader range of brain activity datasets from various scanning protocols and demographic groups could enhance the model's robustness.
  • Multi-modal Integration: Extending Brain-JEPA to integrate modalities like MEG and EEG, or even structural MRI, could provide a comprehensive understanding of brain dynamics.
  • Interpretability: Fine-grained interpretations at the ROI and temporal levels can yield deeper insights into brain functions and disorders.

Conclusion

Brain-JEPA represents a significant advancement in the intersection of AI and neuroscience. Its innovative techniques in Brain Gradient Positioning and Spatiotemporal Masking pave the way for more robust and generalizable brain activity models. This foundation model not only sets new performance benchmarks across various datasets and tasks but also provides valuable insights for future research directions in brain dynamics and AI applications in neuroscience. The paper's contributions underscore a pivotal step towards more accurate and inclusive brain activity analysis, with broad implications for both theoretical neuroscience and practical healthcare solutions.