Residual D²NN: training diffractive deep neural networks via learnable light shortcuts

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Machine Learning

Background:

• Diffractive deep neural networks (D2NNs) are crucial for all-optical machine learning tasks.
• Training deeper D2NNs is challenging due to the gradient vanishing problem.

Purpose of the Study:

• To introduce residual D2NNs (Res-D2NNs) to enable training of substantially deeper diffractive networks.
• To address the gradient vanishing issue in deep diffractive neural networks.

Main Methods:

• Developed diffractive residual learning blocks to learn residual mapping functions.
• Incorporated a learnable light shortcut for direct input-output connection between optical layers.
• Enabled direct gradient backpropagation path for training.

Main Results:

• Res-D2NNs successfully alleviate the gradient vanishing issue in very deep diffractive networks.
• Experimental results show superior performance of Res-D2NNs compared to plain D2NNs.
• Demonstrated effectiveness on image classification and pixel super-resolution tasks.

Conclusions:

• Res-D2NNs represent a significant advancement in training deep all-optical neural networks.
• The proposed residual learning approach enhances the capability and complexity of diffractive AI systems.