Phase-shift determination for a 4  ×  4 intelligent photonic neural network with compatible learning.

Area of Science:

• Photonics
• Artificial Intelligence
• Optical Computing

Background:

• Intelligent photonic circuits (IPCs) are key for advanced optical computing.
• Photonic Mach-Zehnder Interferometers (MZIs) are fundamental components in IPCs.
• Implementing complex-valued neural networks requires precise control over photonic components.

Purpose of the Study:

• To enable photonic intelligent matrices for task-oriented topologies using optimized phase-shift vectors.
• To develop and validate a robust phase determination system for IPCs.
• To demonstrate the proof of concept for phase-shift determination in a 4x4 intelligent photonic neural network.

Main Methods:

• Utilizing photonic Mach-Zehnder Interferometers (MZIs) for matrix representation.
• Employing compatible learning for complex-valued neural networks with digital weight matrices.
• Formulating a phase determination system using nonlinear least squares to refine phase-shift solutions.

Main Results:

• A method for broadcasting MZI matrix representations into arbitrary matrices with optimized phase-shift vectors was developed.
• The robustness of the phase determination system for photonic neural networks was analyzed.
• Numerical experiments with a 4x4 intelligent photonic neural network verified the phase-shift determination concept.

Conclusions:

• Optimized phase-shift vectors are crucial for realizing effective intelligent photonic circuits.
• The developed phase determination system offers a robust approach for photonic neural network implementation.
• This work lays the foundation for advanced, multi-layered photonic neural networks.