An All-Optical Artificial Neural Network


Scatterings image

Researchers have demonstrated a two-layer, all-optical artificial neural network with nonlinear activation functions, which are required to perform complex tasks such as pattern recognition. [Image: Olivia Wang, Peng Cheng Laboratory]

Since the dawn of digital computing, scientists have dreamed of building artificial neural networks that would function like biological brains and solve difficult problems. As nanophotonic circuits became a reality, researchers have tried to create neural networks that would run at the speed of light, but translating a key mathematical component of artificial neurons, the nonlinear activation function, from the electronic to the optical realm proved elusive.

Now, a nine-member team from the Hong Kong University of Science and Technology, Hong Kong, has built an all-optical artificial neural network and applied it to a complex simulation (Optica, doi:10.1364/OPTICA.6.001132). The network incorporates both linear functions, powered by spatial light modulators, and nonlinear activation functions based on the quantum interference effect known as electromagnetically induced transparency (EIT).

Inputs and outputs

Artificial neural networks typically consist of interconnected layers, with the outputs of one layer becoming the inputs of the next layer. The idea, of course, is to simulate the complex interconnections of biological neurons and axons. For the linear functions of their network, the Hong Kong scientists sent an incident coupling laser beam through two spatial light modulators, a Fourier lens and a flip mirror. The programmable Fourier lens manipulated the laser signals.

Earlier attempts to create optical activation functions required high-powered lasers for the nonlinear optical components. This time around, the Hong Kong team generated these activation functions from laser-cooled rubidium-85 atoms, held in a magneto-optical trap between low-powered coupling and probe laser beams. The researchers could produce different activation functions by changing the positions of the counterpropagating beams.

The “Ising” on the cake

To apply the technology to an actual computational problem, the researchers sent signals from the first spatial light modulator into a polarizing beam splitter, where they interacted with nonlinear light signals from the optical trap. The final beams hit a second spatial light modulator before they were recorded by a camera. As a proof of concept, the team simulated a two-dimensional statistical Ising model on a square lattice, a problem from condensed-matter physics.

The scientists say that their all-optical neural network could be scaled up to tackle other, more complex problems—from image recognition to accelerated Monte Carlo simulations.



Publish Date:

Add a Comment

Article Tools