Model-free optimization and parallel architecture towards monolithic-hybrid-photonic-electronic reservoir computing

Area of Science:

• Neuromorphic Computing
• Photonics
• Machine Learning

Background:

• Physical reservoir computing (PRC) leverages nonlinear physical systems for machine learning.
• Monolithic-hybrid-photonic-electronic reservoir computing (MHPE RC) combines photonic nonlinearity with electronic tunability.

Purpose of the Study:

• To analyze the performance of photonic waveguide meshes (WGMs) in MHPE RC.
• To develop and validate an optimization algorithm for enhanced MHPE RC performance.

Main Methods:

• Numerical demonstration of parallel WGM architecture.
• Development of the Lyapunov filtered-minimal redundancy maximal relevance (Lf-mRMR) algorithm for parameter optimization.
• Experimental validation using on-chip silicon photonics.

Main Results:

• The parallel WGM architecture shows efficiency and performance superiority.
• The Lf-mRMR algorithm improves MHPE RC performance, tolerates fabrication errors, and reduces computational complexity.
• The selective parallel architecture for reservoir computing (SPARC) achieves performance comparable to convolutional neural networks.

Conclusions:

• The Lf-mRMR algorithm significantly enhances MHPE RC performance.
• On-chip silicon photonics successfully validates the advantageous performance of Lf-mRMR-assisted RC.
• This approach offers a computationally efficient and robust method for advanced machine learning tasks.