A myriad of devices equipped with artificial intelligence and connected via ultra-fast mobile networks: thanks to these technologies, which are now commonly used, we generate an immense amount of data at any given moment, to be processed immediately and energy efficiently. In order to process this growing amount of information, modern computing systems require complex training or learning phases, which have a considerable impact on energy and the environment. It is estimated that data centre energy consumption will soon account for more than 10% of global electricity use. Training an artificial intelligence system to tell us the best route to take, for example, will pollute more than the car we use to drive it. The development of innovative processors and energy-efficient computing systems is therefore increasingly vital.

Today's most promising way forward for future hardware systems is photonics, which aims to replace electrical circuits with optical technologies that process information in parallel, at very high frequencies, and with minimal energy consumption. However, photonic computing networks require elaborate fabrication and control systems, making large-scale implementation difficult, and their training is as expensive as that of electronic processors.

It is in this context that the team of researchers coordinated by Claudio Conti of the Department of Physics at Sapienza University of Rome, in collaboration with the CNR's Institute of Complex Systems, Cref and the University of Ottawa, has developed an innovative computing device that uses photonics in a completely sustainable way to perform artificial intelligence operations.

This green photonic processor, presented in the journal Photonics Research, uses optical propagation in the air as a virtual network capable of performing complex operations, such as reading manually written characters, which requires the processing of tens of thousands of handwriting styles. The device's artificial intelligence is based on the paradigm of extreme learning machines, which mimic the brain's learning systems and are easy to train. Using only a very low-power laser beam on which the data to be processed are encoded, the photonic device requires no nanofabrication and performs classification and recognition tasks with minimal energy consumption.

It was soon realised that it was unnecessary to engineer complex optical components to create a photonic system capable of learning, but that it was sufficient to exploit the natural complexity that optical field waves have in propagating.

"Our results show that fabricated optical networks, or complex physical processes and materials, are not essential ingredients to perform effective machine learning on an optical setup," says Claudio Conti, "and that all factors essential for learning can be included in optical propagation, through encoding and decoding methods."

Davide Pierangeli, the ISC-CNR researcher who implemented the device, continues: "Our scheme is particularly promising for applications in edge computing - a frontier field in the development of smart sensors that can operate autonomously in dynamic environments."

This study opens up interesting prototype applications, including autonomous driving systems and real-time video recognition.

References:

Photonic extreme learning machine by free-space optical propagation - Davide Pierangeli, Giulia Marcucci, Claudio Conti. - Photonics Research 2021, 9, 1446. DOI: https://doi.org/10.1364/PRJ.423531

Further Information

Claudio Conti
Department of Physics 
claudio.conti@uniroma1.it