Reproduce “Shared Independent Component Analysis for Multi-Subject Neuroimaging” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hugo Richard, Pierre Ablin, Bertrand Thirion, Alexandre Gramfort, Aapo Hyvärinen
ShICA models each subject’s observations as a linear mixture of shared independent components plus subject-specific Gaussian noise, allowing non-orthogonal maps and differing noise. Its estimators recover shared components, with ShICA-ML also using likelihood information. In held-out-subject fMRI reconstruction, ShICA-ML scored 0.200 [0.187, 0.213] on Forrest with 20 components, versus 0.191 [0.177, 0.204] for MVICA; this comparison applies only to the tested setup.
Key insights
- Global identifiability of the ShICA model is established, even when some components are Gaussian, provided there is sufficient diversity in noise variances across views, a stronger result than previous local identifiability findings.
- The ShICA-J algorithm effectively mitigates Multiset CCA's sensitivity to sampling noise by using joint diagonalization to refine unmixing matrix estimates, making it more robust.
- ShICA-ML's maximum likelihood framework efficiently leverages both noise diversity and non-Gaussianity of components, enabling accurate inference for a broader range of data characteristics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.