Lightweight optical neural network based on micro-ring resonator.

Area of Science:

• Photonics and Optical Computing
• Artificial Intelligence and Machine Learning
• Integrated Optics and Nanophotonics

Background:

• Traditional optical neural networks (ONNs) face challenges in efficiency, power consumption, and scalability due to complex architectures.
• Depthwise separable convolution is a key technique for reducing computational complexity in neural networks, but its optical implementation is challenging.
• Micro-ring resonators (MRRs) offer potential for compact and efficient optical processing elements.

Purpose of the Study:

• To propose a novel, lightweight, high-efficiency, and low-power ONN architecture.
• To implement depthwise separable convolution in an optical domain using micro-ring resonators.
• To significantly reduce the resource requirements and energy consumption of ONNs.

Main Methods:

• Introduced the micro-ring-based depthwise separable convolution (MDSC) architecture.
• Utilized add-drop MRRs as convolution kernels for efficient feature extraction in depthwise convolution.
• Employed trans-impedance amplifiers (TIAs) for the pointwise convolution stage, using magnification factors as learnable weights.

Main Results:

• MDSC successfully implements depthwise separable convolution in an optical domain.
• Achieved significant reductions in the number of MRRs, execution time, and energy consumption (up to 3-5 orders of magnitude) compared to traditional MRR-based ONNs.
• Demonstrated scalability and efficiency gains on complex network structures.

Conclusions:

• The proposed MDSC architecture offers a highly efficient and low-power solution for optical neural networks.
• MDSC effectively reduces computational redundancy and hardware complexity in optical deep learning.
• This approach paves the way for practical, large-scale optical AI hardware.