Shared Independent Component Analysis for Multi-Subject Neuroimaging

Authors: Hugo Richard, Pierre Ablin, Bertrand Thirion, Alexandre Gramfort, Aapo Hyvarinen

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show via simulations that Sh ICA-J leads to improved results while being very fast to fit. While Sh ICA-J is based on second-order statistics, we further propose to leverage non-Gaussianity of the components using a maximum-likelihood method, Sh ICA-ML, that is both more accurate and more costly. Further, Sh ICA comes with a principled method for shared components estimation. Finally, we provide empirical evidence on f MRI and MEG datasets that Sh ICA yields more accurate estimation of the components than alternatives.
Researcher Affiliation Academia Hugo Richard Inria Université Paris-Saclay Palaiseau, France Pierre Ablin DMA CNRS and ENS Paris, France Bertrand Thirion Inria Université Paris-Saclay Palaiseau, France Alexandre Gramfort Inria Université Paris-Saclay Palaiseau, France Aapo Hyvärinen Department of Computer Science University of Helsinki Helsinki, Finland
Pseudocode Yes Algorithm 1 Sh ICA-J Input : Covariances Cij = E[xix j ] ( Wi)i Multiset CCA(( Cij)ij) Q Joint Diag(( Wi Cii W i )i) Γij Q Wi Cij W j Q (Φi)i Scaling((Γij)ij) Return : Unmixing matrices (Φi Q Wi)i.
Open Source Code Yes Our code is available at https://github.com/hugorichard/Sh ICA.
Open Datasets Yes f MRI experiments used the following datasets: sherlock [15], forrest [28] , raiders [49] and gallant [49]. In the following experiments we consider the Cam-CAN dataset [56].
Dataset Splits No The paper describes a training and testing split strategy ('The unmixing operators are learned using all subjects and 80% of the runs. Then they are applied on the remaining 20% of the runs using 80% of the subjects...'), but it does not explicitly mention a separate validation set for hyperparameter tuning.
Hardware Specification No Computations were run on a large server using up to 100 GB of RAM and 20 CPUs in parallel. This is not specific enough to identify hardware (e.g., no CPU model, no GPU mentioned).
Software Dependencies No Experiments used Nilearn [3] and MNE [26] for f MRI and MEG data processing respectively, as well as the scientific Python ecosystem: Matplotlib [31], Scikit-learn [45], Numpy [29] and Scipy [61]. We use the Picard algorithm for non-Gaussian ICA [2], and mvlearn for multi-view ICA [47]. No version numbers are provided for these software dependencies.
Experiment Setup Yes In the following synthetic experiments, data are generated according to model (1) with p = 4 components and m = 5 views and mixing matrices are generated by sampling coefficients from a standardized Gaussian. Gaussian components are generated from a standardized Gaussian and their noise has standard deviation Σ 1 2 i (obtained by sampling from a uniform density between 0 and 1) while non-Gaussian components are generated from a Laplace distribution and their noise standard deviations are equal. Data are first reduced using a subject-specific PCA with p = 10 components. The unmixing operators are learned using all subjects and 80% of the runs.