Deeply Learned Robust Matrix Completion for Large-Scale Low-Rank Data Recovery.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Robust Matrix Completion (RMC) is crucial for low-rank data analysis, addressing missing entries and outliers.
• Existing RMC methods often face challenges with scalability and parameter optimization for large datasets.

Purpose of the Study:

• Introduce Learned Robust Matrix Completion (LRMC), a novel, scalable, and learnable non-convex approach for large-scale RMC.
• Develop an effective parameter learning strategy using deep unfolding for optimal LRMC performance.
• Propose a flexible neural network framework to extend deep unfolding for enhanced RMC.

Main Methods:

• Developed a novel non-convex optimization framework for Robust Matrix Completion (RMC).
• Utilized deep unfolding techniques to learn free parameters of the LRMC model.
• Proposed a feedforward-recurrent-mixed neural network for infinite iteration deep unfolding.

Main Results:

• LRMC demonstrates low computational complexity and linear convergence rates.
• Extensive experiments show LRMC outperforms state-of-the-art methods on synthetic and real-world data.
• Validated LRMC in applications like video background subtraction, ultrasound imaging, face modeling, and satellite imagery cloud removal.

Conclusions:

• LRMC provides a superior, efficient, and scalable solution for robust matrix completion.
• Deep unfolding and novel neural network architectures enable effective parameter learning and performance enhancement.
• The proposed LRMC framework is versatile and applicable to diverse real-world data analysis challenges.