Reproduce “Modeling Shared Responses in Neuroimaging Studies through MultiView ICA” 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, Luigi Gresele, Aapo Hyvärinen, Bertrand Thirion, Alexandre Gramfort, Pierre Ablin
MultiView ICA estimates brain responses shared across people while allowing each person a different mapping from shared signals to measurements. It models individual deviations as noise and uses a closed-form likelihood with alternating quasi-Newton optimization. In fMRI time-segment matching on Sherlock, it reaches 0.873 mean accuracy versus 0.607 for CanICA; missing-subject BOLD reconstruction is broadly comparable, not a strict win.
Key insights
- Modeling individual noise directly in the source domain (i.e., `s + n_i`) is key to achieving a tractable, closed-form likelihood function for multi-subject ICA, which was a significant theoretical hurdle previously.
- The model's identifiability and demonstrated robustness to hyperparameter misspecification, such as the noise parameter or exact source density, enhance its reliability and practical applicability in diverse neuroimaging contexts.
- An alternating quasi-Newton optimization approach provides an efficient and robust mechanism for optimizing the non-convex objective, making the statistically rigorous model computationally feasible for large-scale datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.