Scalable Orthonormal Projective NMF via Diversified Stochastic Optimization.

Area of Science:

• Computational neuroscience
• Neuroimaging analysis
• Data-driven discovery science

Background:

• Large-scale neuroimaging initiatives provide opportunities for brain structure and function discovery.
• Orthonormal Projective Non-negative Matrix Factorization (opNMF) is valuable for exploring multivariate relationships in big data.
• Standard opNMF faces scalability challenges in large cohort studies due to computational complexity.

Purpose of the Study:

• To address the computational limitations of opNMF for large-scale neuroimaging data analysis.
• To introduce a novel stochastic optimization approach for opNMF.
• To enhance the scalability and applicability of opNMF in big data research.

Main Methods:

• Implemented a stochastic optimization approach learning over mini-batches of data.
• Utilized repulsive point processes to diversify mini-batches and reduce update variance.
• Validated the framework on gray matter density maps from the OASIS dataset (1000 subjects).

Main Results:

• Demonstrated significant reduction in computational cost using mini-batch operations.
• Confirmed that the novel optimization method does not compromise the accuracy of opNMF factors.
• Showcased maintained interpretability of factors compared to standard opNMF.

Conclusions:

• The proposed stochastic opNMF model enhances scalability for big neuroimaging data.
• This advancement enables new investigations into brain structure in health and disease.
• Improved computational efficiency facilitates broader application of opNMF in neuroscience research.