All-optical diffractive neural networks take any light

December 03, 2019 //By Julien Happich
neural networks
Researchers from the UCLA Engineering Institute for Technology Advancement have demonstrated that trained deep neural networks could be implemented as a set of all-optical diffractive surfaces able to process images at the speed of light, including across multiple frequencies of incoherent light.

A diffractive neural network is first designed in a computer using deep learning techniques and materials modelling techniques. In order to design a broadband diffractive network, the researchers first selected a set of physical parameters to be optimized, such as the thickness of each neuron within the diffractive layers, enabling the control of the phase modulation profile of each diffractive layer in the network.

In addition, to correctly represent the forward model of the broadband light propagation within their optical neural network, the scientists also took into consideration the material dispersion, including the real and imaginary parts of the refractive index of the network material as a function of the wavelength.


The diffractive layers are 3D printed over a surface that is
larger than their active area to avoid bending of the layers.
These extra regions do not modulate the light and are
covered by aluminium, preventing stray light into the system.
The active area of the first diffractive layer is 1×1cm, while
the other layers have active areas of 5×5cm.

In a paper titled "Design of task-specific optical systems using broadband diffractive neural networks" published in the Light: Science & Applications journal, the researchers explain how for each wavelength within the input light spectrum, the all-optical neural network (AONN) yields a unique complex (phase and amplitude) modulation corresponding to the transmission coefficient of each neuron across the different diffractive layers, as determined by its physical thickness obtained during the training of the diffractive optical network.


Prototype of a broadband diffractive deep neural network.
Credit: Ozcan Lab at UCLA.

After training comes the fabrication, the authors used a 3-D printer to create the two diffractive layers corresponding to the trained neurons (at a 0.5mm pitch) and tested the AONN using a pulsed THz source emitting a continuum of wavelengths between 60 and 3,000 micrometers. They reported experimental results in agreement with their corresponding numerical designs. Since the connection between the input and output planes of a diffractive neural network is established via diffraction of light through passive layers, the inference process and the associated optical computation does not consume any power except the light used to illuminate the object of interest.


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.