Scalable NMF via linearly optimized data compression.

Area of Science:

• Neuroimaging
• Clinical Neuroscience
• Computational Neuroscience

Background:

• Orthonormal projective non-negative matrix factorization (opNMF) is valuable for brain representations in health and disease.
• opNMF's constraints yield intuitive, well-localized factors but are computationally expensive for large datasets.

Purpose of the Study:

• To propose scalable optimization schemes for opNMF suitable for large-scale neuroimaging data.
• To overcome the computational limitations of traditional opNMF.

Main Methods:

• Replacing high-dimensional data matrices with Singular Value Decomposition (SVD) and QR decompositions.
• Integrating these decompositions with the opNMF multiplicative update algorithm.
• Developing novel optimization schemes for enhanced computational efficiency.

Main Results:

• Demonstrated significant speed-up in computation time for opNMF.
• Maintained low memory consumption.
• Ensured accuracy of the factorization results was not compromised.

Conclusions:

• The proposed scalable opNMF methods enable efficient analysis of large-scale neuroimaging datasets.
• These advancements facilitate broader application of opNMF in clinical neuroscience.
• The optimized approach balances computational efficiency with analytical accuracy.