Microring-based programmable coherent optical neural networks.

Area of Science:

• Integrated photonics
• Optical computing
• Deep learning hardware

Background:

• Deep learning tasks rely heavily on matrix multiplication and nonlinear activation.
• Integrated photonics circuits offer potential for specialized hardware accelerators.
• Existing solutions face challenges in footprint and energy efficiency.

Purpose of the Study:

• To design, simulate, and train a novel optical neural network (ONN).
• To leverage microring resonators for enhanced device footprint and energy efficiency.
• To develop efficient training algorithms for optical components.

Main Methods:

• Utilized tunable coupled double ring structures for linear multiplication layers.
• Employed modulated microring resonators for reconfigurable nonlinear activation.
• Developed optimization algorithms using transfer matrix method and automatic differentiation for parameter tuning (e.g., applied voltages).

Main Results:

• Successfully designed, simulated, and trained an ONN based entirely on microring resonators.
• Demonstrated advantages in device footprint and energy efficiency compared to conventional accelerators.
• Validated the effectiveness of the developed optimization algorithms for training optical components.

Conclusions:

• Microring resonator-based ONNs are a promising approach for efficient deep learning acceleration.
• The proposed architecture offers significant improvements in size and power consumption.
• Advanced optimization techniques enable effective training of complex optical neural networks.