braindyn-sheaf-neural-ode

name: braindyn-sheaf-neural-ode description: "BrainDyn: A Sheaf Neural ODE framework for modeling continuous-time brain dynamics on structured graphs. Combines LSTM stalks with sheaf Laplacian message passing and neural ODE evolution. Apply when: brain dynamics modeling, fMRI/EEG forecasting, generative brain models, neural ODEs, sheaf theory, brain graph networks, perturbation prediction, synthetic brain data. Keywords: sheaf neural ODE, brain dynamics, fMRI modeling, EEG forecasting, brain graphs, sheaf Laplacian, neural ODE, brain regions, generative dynamics."

BrainDyn: Sheaf Neural ODE for Generative Brain Dynamics

A novel framework combining sheaf theory, neural ODEs, and structured brain graphs for continuous-time brain dynamics modeling.

Overview

BrainDyn addresses the gap between LLM/RNNs (ignore anatomical organization) and simple graph networks (insufficient expressiveness for brain-like dynamics) by introducing:

1. Sheaf Neural ODE: Continuous-time dynamics on structured brain graphs
2. LSTM Stalks: Encode recent activity history per brain region
3. Sheaf Laplacian: Facilitate message passing between neuronal units

Core Architecture

Temporal Encoding

• LSTM over Sliding Windows: Produce hidden states (stalks) per brain region
• Restriction Maps: Project stalks through learnable maps to edge-specific shared spaces
• Activity History: Capture recent dynamics for each region

Message Passing

• Sheaf Laplacian: Characterize discrepancies between neighboring nodes in shared spaces
• Edge-Specific Spaces: Different spaces for different connections
• Expressive Dynamics: Beyond simple message passing rules

Continuous Evolution

• Neural ODE: Govern continuous-time evolution of neuronal activity
• Smooth Dynamics: Natural temporal progression
• Brain-Region Aligned: Components align with anatomical organization

Key Innovations

1. Sheaf Theory Integration: Mathematical framework for multi-space consistency
2. Anatomical Alignment: Brain region correspondence maintained
3. Continuous-Time Modeling: Natural temporal dynamics (vs discrete steps)
4. Multi-Modality Support: fMRI, EEG, simulated spiking data

Applications

Data Modalities

• Resting-state fMRI: PNC dataset
• Scalp EEG: Focal epilepsy (TUSZ dataset)
• Spiking Network: NEST simulator output

Downstream Tasks

• Brain activity forecasting
• In silico perturbation prediction
• Synthetic brain data generation
• Brain transient analysis
• Dynamics inference

Implementation

Model Components

# Conceptual architecture
class BrainDyn:
    - LSTMEncoder: Per-region activity encoding
    - RestrictionMaps: Learnable stalk projections
    - SheafLaplacian: Multi-space message passing
    - NeuralODE: Continuous dynamics solver

Training Approach

• Supervised forecasting across modalities
• Representation learning for downstream tasks
• Perturbation prediction evaluation

Mathematical Framework

Sheaf Structure

• Stalks: Hidden states per brain region (LSTM outputs)
• Restriction Maps: Learnable transformations to edge spaces
• Sheaf Laplacian: Aggregates discrepancies across edges
• Neural ODE: dx/dt = f(x, L_sheaf, θ)

Expressiveness

• Captures region-specific dynamics
• Maintains anatomical structure
• Enables complex temporal patterns
• Supports perturbation analysis

Biological Validity

• Brain region alignment
• Multi-modal applicability (fMRI, EEG, spikes)
• Continuous temporal dynamics
• Perturbation prediction capability

Reference

Paper: "BrainDyn: A Sheaf Neural ODE for Generative Brain Dynamics" arXiv ID: 2605.19324 Authors: Siddharth Viswanath, Panayiotis Ketonis, Chen Liu, Michael Perlmutter, Dhananjay Bhaskar, Smita Krishnaswamy Published: 2026-05-19 Category: cs.LG (Machine Learning)

Activation Keywords

sheaf neural ODE, brain dynamics, fMRI modeling, EEG forecasting, brain graphs, sheaf Laplacian, neural ODE, brain regions, generative dynamics, brain transients, perturbation prediction, LSTM stalks, continuous dynamics, anatomical alignment