A Two-Timescale Duplex Neurodynamic Approach to Biconvex Optimization.

Area of Science:

• Computational Neuroscience
• Optimization Theory
• Machine Learning

Background:

• Biconvex optimization problems are prevalent in various scientific domains.
• Traditional optimization methods can suffer from instability and local minima.
• Recurrent Neural Networks (RNNs) offer potential for dynamic optimization but require careful design.

Purpose of the Study:

• To develop a novel neurodynamic system for solving constrained biconvex optimization problems.
• To enhance stability and global convergence properties compared to existing methods.
• To apply the proposed system to real-world problems like L1-constrained nonnegative matrix factorization.

Main Methods:

• A two-timescale duplex neurodynamic system comprising two collaborative RNNs.
• Utilizing two distinct timescales to mitigate system instability.
• Integrating Particle Swarm Optimization (PSO) to refine RNN initial states and escape local minima.

Main Results:

• The proposed system demonstrates global convergence to the global optimum with probability one.
• Empirical validation through benchmark problems confirms system performance.
• Successful application to L1-constrained nonnegative matrix factorization showcases practical utility.

Conclusions:

• The two-timescale duplex neurodynamic system provides a robust and stable approach for constrained biconvex optimization.
• The integration of PSO effectively addresses the challenge of local minima.
• This neurodynamic framework offers a promising tool for advanced machine learning and signal processing tasks.