Title
Reservoir Brains: Recurrent Machines for Dynamic Brain Signals
Short description and the goals for the OHBM BrainHack
This project explores lightweight recurrent neural networks and state-space models, including echo state networks and connectome-informed architectures, for modeling and decoding fMRI time series. We aim to determine whether these models can capture meaningful brain dynamics and improve tasks such as behavioral prediction, clinical classification, stimulus decoding, and forecasting future brain activity.
Using open datasets such as the Human Connectome Project and Courtois Neuromod, we will build an open and reproducible pipeline to compare recurrent approaches against standard baselines based on static functional connectivity and classical machine learning. A key exploratory direction is to investigate whether human functional connectomes can guide recurrent model architectures and whether individual differences in cognition are reflected in the computational properties of these brain-informed models.
The project welcomes contributors interested in neuroimaging, machine learning, time-series analysis, and computational neuroscience. No prior experience with recurrent models is required.
All welcome to join us!
Link to the Project
https://github.com/ReservoirBrains
Image/Logo for the OHBM brainhack website
https://raw.githubusercontent.com/ReservoirBrains/RB-DMT/main/logo.png
Project lead
Xavier Hinaut — GitHub: neuronalX — Discord: @Xaviolo
Yannis Bendi-Ouis — GitHub: Naowak— Discord: naowak_
Clara Moreau
Clara El Khantour — GitHub: claraelk — Discord: clacla2410
Mathieu Dugré — GitHub: mathdugre — Discord: na_math
Achille Gillig — GitHub: Achillegillig — Discord: ach_etyl
Main Hub
Bordeaux
Link to the Project pitch
https://docs.google.com/presentation/d/1cqt38GI3ni4E6_nxoX_u7_kgO0HHWx1Qipj0rAMevOs/edit?usp=sharing
Other hubs covered by the leaders
Skills
Python
Neuroimaging / fMRI analysis
Resting-state/task fMRI
Functional connectivity, Connectome modeling
Individual differences, Behavioral prediction
Machine learning
Recurrent neural networks
State Space Models
Reproducible research
Recommended tutorials for new contributors
Human Connectome Project data overview: https://www.humanconnectome.org/study/hcp-young-adult
Courtois Neuromod documentation: https://docs.cneuromod.ca/en/latest/
Courtois Neuromod processed derivatives: https://github.com/courtois-neuromod/cneuromod.processed
OpenNeuro TCP dataset: https://openneuro.org/datasets/ds005237/versions/1.1.3
Nilearn tutorials: https://nilearn.github.io/stable/auto_examples/index.html
scikit-learn getting started: https://scikit-learn.org/stable/getting_started.html
ReservoirPy documentation: https://reservoirpy.readthedocs.io/
Good first issues
Set up the project repository structure and README
Create an environment file with the required Python dependencies
Install and test ReservoirPy with a minimal example
Compare candidate datasets and document access requirements
Prepare example parcellated fMRI time series using Yeo or Schaefer atlases
Implement baseline models using static functional connectivity
Implement a simple echo state network / recurrent state-space model
Prototype fMRI forecasting at time t + N
Implement a connectome-informed recurrent model using functional connectivity as the transition topology
Test whether subject-specific model quality relates to cognitive or behavioral performance
Create basic plots comparing model predictions and observed outcomes
Document the workflow so new contributors can reproduce and extend it
Twitter summary
Can human connectomes shape better recurrent models? This OHBM BrainHack project will build an open pipeline for fMRI dynamics using echo state networks and state-space models to predict behavior, decode stimuli, forecast brain activity, and test whether individual brain connectivity translates into better or worse model performance.
Short name for the Discord chat channel (~15 chars)
ReservoirBrains
Please read and follow the OHBM Code of Conduct
Title
Reservoir Brains: Recurrent Machines for Dynamic Brain Signals
Short description and the goals for the OHBM BrainHack
This project explores lightweight recurrent neural networks and state-space models, including echo state networks and connectome-informed architectures, for modeling and decoding fMRI time series. We aim to determine whether these models can capture meaningful brain dynamics and improve tasks such as behavioral prediction, clinical classification, stimulus decoding, and forecasting future brain activity.
Using open datasets such as the Human Connectome Project and Courtois Neuromod, we will build an open and reproducible pipeline to compare recurrent approaches against standard baselines based on static functional connectivity and classical machine learning. A key exploratory direction is to investigate whether human functional connectomes can guide recurrent model architectures and whether individual differences in cognition are reflected in the computational properties of these brain-informed models.
The project welcomes contributors interested in neuroimaging, machine learning, time-series analysis, and computational neuroscience. No prior experience with recurrent models is required.
All welcome to join us!
Link to the Project
https://github.com/ReservoirBrains
Image/Logo for the OHBM brainhack website
https://raw.githubusercontent.com/ReservoirBrains/RB-DMT/main/logo.png
Project lead
Xavier Hinaut — GitHub: neuronalX — Discord: @Xaviolo
Yannis Bendi-Ouis — GitHub: Naowak— Discord: naowak_
Clara Moreau
Clara El Khantour — GitHub: claraelk — Discord: clacla2410
Mathieu Dugré — GitHub: mathdugre — Discord: na_math
Achille Gillig — GitHub: Achillegillig — Discord: ach_etyl
Main Hub
Bordeaux
Link to the Project pitch
https://docs.google.com/presentation/d/1cqt38GI3ni4E6_nxoX_u7_kgO0HHWx1Qipj0rAMevOs/edit?usp=sharing
Other hubs covered by the leaders
Skills
Python
Neuroimaging / fMRI analysis
Resting-state/task fMRI
Functional connectivity, Connectome modeling
Individual differences, Behavioral prediction
Machine learning
Recurrent neural networks
State Space Models
Reproducible research
Recommended tutorials for new contributors
Human Connectome Project data overview: https://www.humanconnectome.org/study/hcp-young-adult
Courtois Neuromod documentation: https://docs.cneuromod.ca/en/latest/
Courtois Neuromod processed derivatives: https://github.com/courtois-neuromod/cneuromod.processed
OpenNeuro TCP dataset: https://openneuro.org/datasets/ds005237/versions/1.1.3
Nilearn tutorials: https://nilearn.github.io/stable/auto_examples/index.html
scikit-learn getting started: https://scikit-learn.org/stable/getting_started.html
ReservoirPy documentation: https://reservoirpy.readthedocs.io/
Good first issues
Set up the project repository structure and README
Create an environment file with the required Python dependencies
Install and test ReservoirPy with a minimal example
Compare candidate datasets and document access requirements
Prepare example parcellated fMRI time series using Yeo or Schaefer atlases
Implement baseline models using static functional connectivity
Implement a simple echo state network / recurrent state-space model
Prototype fMRI forecasting at time t + N
Implement a connectome-informed recurrent model using functional connectivity as the transition topology
Test whether subject-specific model quality relates to cognitive or behavioral performance
Create basic plots comparing model predictions and observed outcomes
Document the workflow so new contributors can reproduce and extend it
Twitter summary
Can human connectomes shape better recurrent models? This OHBM BrainHack project will build an open pipeline for fMRI dynamics using echo state networks and state-space models to predict behavior, decode stimuli, forecast brain activity, and test whether individual brain connectivity translates into better or worse model performance.
Short name for the Discord chat channel (~15 chars)
ReservoirBrains
Please read and follow the OHBM Code of Conduct