Multiscale diffractive U-Net: a robust all-optical deep learning framework modeled with sampling and skip...

Area of Science:

• Optical Computing
• Deep Learning
• Artificial Intelligence

Background:

• Diffractive deep neural networks (D2NNs) offer advantages in speed, throughput, and energy efficiency for optical learning.
• Challenges in D2NN development include network depth limitations and layer misalignment.

Purpose of the Study:

• To propose a robust all-optical network framework, the multiscale diffractive U-Net (MDUNet), to overcome D2NN limitations.
• To improve depth expansion and alignment robustness in all-optical networks.

Main Methods:

• Introduced a multiscale diffractive U-Net (MDUNet) architecture.
• Incorporated sampling and skip connections to enhance network depth and alignment robustness.
• Utilized ensemble learning for optoelectronic hybrid network construction.

Main Results:

• MDUNet achieved high accuracy: 98.81% on MNIST and 89.11% on Fashion-MNIST.
• Further improved accuracy to 99.06% (MNIST) and 89.86% (Fashion-MNIST) with ensemble learning.
• Demonstrated significant improvements in depth expansion and alignment robustness compared to existing frameworks.

Conclusions:

• The proposed MDUNet framework effectively addresses key limitations in diffractive deep neural networks.
• MDUNet represents a significant advancement in all-optical learning frameworks, achieving state-of-the-art accuracy.
• Optoelectronic hybrid networks built with MDUNet show potential for further performance gains.