From combinatorial optimization algorithms to artificial neural networks (NN), modern digital computers have undergone incredible development in recent decades. Yet, they struggle to keep up with the growing trend of high-density applications.
Artificial intelligence, in particular machine learning, requires the ability to process increasingly large and complex quantities of data.

One of the biggest obstacles lies in the architecture of digital computers, known as the Von Neumann architecture, named after its creator. In the Von Neumann model, the memory and the processor are physically separate, so information processing requires a continuous passage between them, causing latency times in calculation and extra energy consumption.

Even in parallel computation using multiple processors, e.g., Graphic Processor Units, the calculation architecture remains similar, and information processing suffers from the same Von Neumann “bottleneck.”

A fundamentally different path is one that tries to reproduce information processing by taking inspiration from the brain’s functioning. This path brings the memory or signal as close as possible to the processing units, thus obtaining advantages in terms of speed and efficiency.

Photonic computers offer interesting prospects in this direction. Using materials with specific optical qualities, they can manipulate the waves of light passing through them to perform various mathematical calculations. In this way, they are able to reproduce the neuronal structure of the brain by managing high-dimensional inputs, such as images, in which a lot of information arrives at the same time—and to process them in parallel using a memory distributed along the network.

To understand how it works, let’s imagine a library that has the books scattered along the corridors instead of having the books to consult crammed into a single place. This allows you not to have to travel long distances to consult them each time but to be able to access the shelves closest to where you are from time to time.

Up to this point, the most interesting results have been with electro-optical architectures, where a combination of optics and electronics is used. The challenge now is to implement a purely optical system capable of performing linear and non-linear operations, an essential requirement for creating “deep” photonic neural networks, i.e., with multiple levels of computation.

A recent work, “Large-scale Photonic Computing with Nonlinear Disordered Media,” published in Nature Computational Science, offers an interesting perspective on the type of material to be used.

The research results from an international collaboration between research groups from the Sorbonne University in Paris, the ETH in Zurich, the Tsinghua University in Beijing and the Enrico Fermi Research Center.

The researchers experimented with a micrometric plate of lithium niobate obtained by assembling nanoparticles of the material in a disordered manner with controlled thickness. These materials have a polycrystalline structure with a high second-order nonlinearity coefficient and can generate nonlinear light with multiple scattering due to their composition.

« In practice, they create multiple paths for the light that form a network of connections and simultaneously change the color of the light that passes through them,» said Romolo Savo of CREF. «These two phenomena allow us to perform some of the mathematical operations typical of neural networks in an analogue way, with the great advantage of being able to increase the amount of information processed without increasing the energy cost of the process».

This type of photonic computer could allow the development of artificial intelligence systems with greater calculation speed, energy efficiency and lower environmental impact.

Contacts:

romolo.savo@cref.it