Co-designed metaoptoelectronic deep learning.

Area of Science:

• Optics and Photonics
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models, particularly deep neural networks (DNNs), excel at image recognition tasks.
• Traditional DNN implementation relies heavily on digital hardware, which can be energy-intensive and computationally demanding.
• Integrating optical components offers a potential pathway to accelerate DNN computations and reduce power consumption.

Purpose of the Study:

• To co-design a meta-optical system with electronic hardware for efficient deep learning image recognition.
• To demonstrate the feasibility of using optical convolution blocks for DNN layers.
• To evaluate the performance of the hybrid optical-digital system in an image classification task.

Main Methods:

• A meta-optical system featuring a reflective metasurface was designed to perform optical convolution for one DNN layer.
• Electronic hardware was integrated with the optical component for a hybrid deep learning architecture.
• The combined optical and digital components were jointly optimized for image classification.

Main Results:

• The meta-optical system achieved 65% accuracy in the image classification task.
• This accuracy is comparable to a fully-digital network's performance (66% accuracy).
• The results validate the potential of hybrid optical-digital approaches for DNN acceleration.

Conclusions:

• Co-designing meta-optical systems with electronic hardware is a viable strategy for deep learning image recognition.
• Optical convolution blocks can effectively replace digital layers in DNNs, offering comparable performance.
• This hybrid approach presents a promising direction for developing faster and more energy-efficient AI hardware.