Utilizing Deep Learning Neural Networks for Optical Signal Processing

2019-12-04

Diffraction deep neural network is an optical machine learning framework that integrates deep learning, optical diffraction, and the interaction between light and matter, to design diffraction surfaces that can perform optical computations at the speed of light. First, deep learning techniques are used to design diffraction neural networks in computers, and then physical manufacturing of the neural network's design layers is carried out using technologies such as 3D printing or lithography. Since the connection between the input and output planes of the diffraction neural network is established through the diffraction of light passing through a passive layer, the inference process and related optical computations do not consume any power, except for the light used to illuminate specific objects.


The diffraction optical network developed by researchers at the University of California, Los Angeles provides a low-power, low-latency, and highly scalable machine learning platform that can be applied in various fields such as robotics, autonomous driving, and defense industry. In addition to providing statistical inference and generalization of data categories, diffraction neural networks have also been used to design deterministic optical systems, such as thin imaging systems.Generally, the development of diffraction network models is aimed at processing information of a single wavelength, thus requiring monochromatic and coherent illumination sources, which differ from incoherent and continuous-wavelength ambient light, making it broadband. To overcome this limitation, researchers at the University of California, Los Angeles have designed diffraction networks that can process information using continuous wavelengths, thereby extending this all-optical computing framework to broadband optical signals. They published their research results in "Light: Science & Applications," demonstrating the success of this new framework by creating a series of optical components that filter broadband input light into desired subbands. These deep learning diffraction systems also control the precise position of each filtered band radiated on the output plane, demonstrating wavelength demultiplexing of the terahertz (THz) part of the electromagnetic spectrum through spatial control.


After being designed in computers, these broadband diffraction networks are manufactured using 3D printers and tested using pulsed THz light sources that emit continuous wavelengths between 60 and 3000 micrometers. The experimental results obtained with these 3D-printed diffraction networks closely match their corresponding numerical designs, highlighting the experimental robustness of broadband diffraction optical networks.Simultaneous analysis and processing of light with multiple wavelengths provide opportunities to enhance the inference and generalization capabilities of diffraction optical networks, enabling them to perform machine learning tasks such as all-optical object recognition, as well as design deterministic and task-specific optical components, thereby expanding the optical design space beyond human intuition.


This new approach is also widely applicable to different parts of the electromagnetic spectrum, including the visible light range, making it an important milestone for diffraction optical networks to be widely used in modern optical components and machine learning systems, playing a role in fields such as robotics, autonomous driving, and surveillance.

RECENT AI NEWS

RECENT AI TOOLS