Large-Scale Independent Vector Analysis (IVA-G) via Coresets

Area of Science:

• Signal Processing
• Machine Learning
• Neuroimaging Analysis

Background:

• Joint blind source separation (JBSS) analyzes multiple datasets by factorizing them into statistically dependent sources.
• Existing JBSS methods face computational challenges, limiting their application to large numbers of datasets.

Purpose of the Study:

• To develop an efficient JBSS methodology applicable to a large number of datasets.
• To improve the scalability and generalizability of JBSS techniques.

Main Methods:

• Proposed a coreset selection method to identify a representative subset of datasets for efficient JBSS.
• Investigated two JBSS methods: an extension of independent vector analysis with a Gaussian model (IVA-G) and generalized joint diagonalization (GJD).
• Derived nonidentifiability conditions and applied the coreset method to enhance generalizability.

Main Results:

• The proposed 'coreIVA-G' method demonstrated significant scalability advantages over existing JBSS methods.
• Achieved superior source separation performance on simulated and real functional magnetic resonance imaging (fMRI) data.
• The coreset approach effectively minimized discrepancy between subset and full dataset statistics.

Conclusions:

• Efficient JBSS is achievable by utilizing a representative data subset (coreset).
• The coreIVA-G method offers a scalable and effective solution for analyzing large-scale multi-dataset problems, particularly in neuroimaging.
• This approach overcomes the computational intractability of traditional JBSS methods for numerous datasets.