Partitionable High-Efficiency Multilayer Diffractive Optical Neural Network.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Machine Learning

Background:

• Multilayer diffractive optical neural networks (DIFF-ONNs) face challenges in system setup and flexibility for layer/data size changes.
• Traditional DIFF-ONNs suffer from increasing device count and exponential energy loss with more layers.

Purpose of the Study:

• To develop a partitionable adaptive multilayer DIFF-ONN addressing setup complexity and flexibility limitations.
• To enable rapid and flexible construction of DIFF-ONNs without optical path readjustment.
• To mitigate energy loss issues inherent in deep multilayer optical networks.

Main Methods:

• A novel architecture for partitionable adaptive multilayer diffractive optical neural networks.
• Strategic partitioning of diffractive devices to enable modular construction.
• Experimental evaluation using MNIST and MNIST fashion databases for classification tasks.

Main Results:

• Achieved 89.1% accuracy on the MNIST database.
• Achieved 81.0% accuracy on the MNIST fashion database.
• Demonstrated state-of-the-art classification performance for the proposed optical neural network.

Conclusions:

• The partitionable adaptive multilayer DIFF-ONN architecture effectively resolves setup issues and enhances flexibility.
• The proposed system avoids linear scaling of optical devices and exponential energy decay.
• This adaptable architecture can be extended for diverse diffraction devices and spectral bands, reaching cutting-edge performance.