Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Riemannian Flow Matching for Brain Connectivity Matrices via Pullback Geometry
Authors: Antoine Collas, Ce Ju, Nicolas Salvy, Bertrand Thirion
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate DIFFEOCFM on three large-scale f MRI datasets with more than 4600 scans from 2800 subjects (ADNI, ABIDE, OASIS-3) and two EEG motor imagery datasets with over 30000 trials from 26 subjects (BNCI2014-002 and BNCI2015-001). It enables fast training and achieves stateof-the-art performance, all while preserving manifold constraints. |
| Researcher Affiliation | Academia | Antoine Collas , Ce Ju, Nicolas Salvy, Bertrand Thirion Inria, CEA, Université Paris-Saclay Palaiseau, France EMAIL |
| Pseudocode | Yes | Algorithm 1: DIFFEOCFM: Training Input: step size h; samplers πY, ϕ#p, ϕ#q Output: Trained parameters θ Initialize θ; while not converged do Sample y πY, t U([0, 1]); Sample z0 ϕ#p( | y), z1 ϕ#q( | y); L u E θ (t, (1 t)z0 +tz1, y) (z1 z0) 2 E; θ optimizer-step(L); end Algorithm 2: DIFFEOCFM: Sampling Input: label y; steps L; step size h; trained θ Output: Generated sample x Sample z0 ϕ#p( | y); for ℓ= 0 to L 1 do zℓ+1 Runge-Kutta-step(u E θ , zℓ, y, h); end x ϕ 1(z L); |
| Open Source Code | Yes | Code: https://github.com/antoinecollas/DiffeoCFM |
| Open Datasets | Yes | We evaluate DIFFEOCFM on three large-scale f MRI datasets with more than 4600 scans from 2800 subjects (ADNI, ABIDE, OASIS-3) and two EEG motor imagery datasets with over 30000 trials from 26 subjects (BNCI2014-002 and BNCI2015-001). ... We use three publicly available resting-state f MRI datasets. The ABIDE dataset [ 53 ]... The ADNI dataset [ 73 ]... The OASIS-3 dataset [ 42 ]... We use two publicly available EEG motor imagery datasets from the BCI competition. The BNCI2014-002 dataset [ 67 ]... The BNCI2015-001 dataset [ 25 ]... |
| Dataset Splits | Yes | We report mean and standard deviations computed across 10 random train-test splits with subject-level grouping, ensuring that scans from the same subject do not appear in both training and test sets. ... We report performance on a leave-one-session-out protocol on BNCI2014-002 and on cross-session experiments on BNCI2015-001. The reported standard deviations are computed over 5 inner splits and averaged over sessions. |
| Hardware Specification | Yes | All experiments were run within 10 hours on a single Nvidia A40 GPU with a 32-cores cpu. |
| Software Dependencies | No | Numerical computation was enabled by the scientific Python ecosystem: Matplotlib [ 33 ], Scikit-learn [ 61 ], Numpy [ 30 ], Scipy [ 72 ], Py Torch [ 59 ], f MRIprep [ 24 ], Nilearn [ 14 ], joblib [ 36 ], Py Riemann [ 5 ], torchcfm [ 69 ], torchdiffeq [ 11 ], pandas [ 58 ], and moabb [ 34 ]. |
| Experiment Setup | Yes | TRIANGDDPM, TRIANGCFM and DIFFEOCFM employ a two-layer MLP with 512 hidden units, trained using Adam W [ 50 ] with a learning rate of 10 3 and batch size 64. Training runs for 200 epochs on f MRI and 2000 epochs on EEG. RIEMCFM has a 6-layer MLP with 512 hidden units trained with Adam W (learning rate of 10 4), as recommended in [ 12 ]. These four methods use the dopri5 method from the torchdiffeq [ 11 ] library for time integration. |