Aligning individual brains with fused unbalanced Gromov Wasserstein
Authors: Alexis Thual, Quang Huy TRAN, Tatiana Zemskova, Nicolas Courty, Rémi Flamary, Stanislas Dehaene, Bertrand Thirion
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We design three experiments to assess the performance of FUGW. In Experiments 1 and 2, we are interested in assessing if aligning pairs of individuals with FUGW increases correlation between subjects compared to a baseline correlation. We also compare the ensuing gains with those obtained when using the competing method MSM [35, 36] to align subjects. In Experiment 3, we derive a barycenter of individuals and assess its ability to capture fine-grained details compared to classical methods. |
| Researcher Affiliation | Academia | Alexis Thual Cognitive Neuroimaging Unit, INSERM, CEA, CNRS, Neuro Spin center, Gif sur Yvette, France Mind, Inria Paris-Saclay, Palaiseau, France Inserm, Collège de France, Paris, France... Huy Tran CMAP, Ecole Polytechnique, Palaiseau, France Université Bretagne-Sud, CNRS, IRISA, Vannes, France... Tatiana Zemskova Mind, Inria Paris-Saclay, Palaiseau, France... Nicolas Courty Université Bretagne-Sud, CNRS, IRISA, Vannes, France... Rémi Flamary CMAP, Ecole Polytechnique, Palaiseau, France... Stanislas Dehaene Cognitive Neuroimaging Unit, INSERM, CEA, CNRS, Neuro Spin center, Gif sur Yvette, France Inserm, Collège de France, Paris, France... Bertrand Thirion Mind, Inria Paris-Saclay, Palaiseau, France |
| Pseudocode | Yes | Details concerning the lower bound as well as the corresponding BCD iteration can be found in the Appendix (see Alg. S1). |
| Open Source Code | Yes | 1https://github.com/alexisthual/fugw provides a Py Torch [28] solver with a scikit-learn [29] compatible API |
| Open Datasets | Yes | In all three experiments, we leverage data from the Individual Brain Charting dataset [31]. It is a longitudinal study on 12 human subjects, comprising 400 f MRI maps per subject collected on a wide variety of stimuli (motor, visual, auditory, theory of mind, language, mathematics, emotions, and more), movie-watching data, T1-weighted maps, as well as other features such as retinotopy which we don t use in this work. We leverage these 400 f MRI maps. The training, validation and test sets respectively comprise 326, 43 and 30 contrast maps acquired for each individual of the dataset. Tasks and MRI sessions differ between each of the sets. More details, including preprocessing, are provided in Supplementary Materials. [31] Ana Luísa Pinho et al. Individual Brain Charting, a high-resolution f MRI dataset for cognitive mapping . In: Scientific Data 5 (June 2018), p. 180105. DOI: 10.1038/sdata.2018.105. URL: https://hal.archives-ouvertes.fr/hal-01817528. |
| Dataset Splits | Yes | The training, validation and test sets respectively comprise 326, 43 and 30 contrast maps acquired for each individual of the dataset. |
| Hardware Specification | Yes | Computation took about 100 hours using 4 Tesla V100-DGXS-32GB GPUs. More precisely, it takes about 4 minutes to compute one coupling between a source and target 10k-vertex hemisphere on a single GPU, when the solver was set to run 10 BCD and 400 Sinkhorn iterations. |
| Software Dependencies | No | The paper mentions using 'Py Torch [28] solver with a scikit-learn [29] compatible API' but does not specify the version numbers for PyTorch or scikit-learn. |
| Experiment Setup | Yes | We set α = 0.5, ρ = 1 and ε = 10 3. Hyper-parameters used to obtain these results were chosen after running a grid search on α, ε and ρ and evaluating it on the validation dataset. ... when the solver was set to run 10 BCD and 400 Sinkhorn iterations. |