In a groundbreaking and game-changing study that could revolutionize the field of artificial intelligence, researchers have successfully demonstrated that photonic chips can be utilized to train neural networks using on-chip backpropagation—the most widely used method for training these complex systems.

This astonishing breakthrough paves the way for the development of futuristic, optically driven, and energy-efficient machine learning technologies, which hold the potential to dramatically reduce the carbon footprint and costs associated with AI computation.

Neural networks, an approach to machine learning conceptually inspired by the miraculous biology of the human brain, have emerged as the cornerstone of many cutting-edge scientific and commercial AI technologies, including the much-talked-about ChatGPT architectures. As neural networks continue to make significant strides and become increasingly ubiquitous, the energy required to power these awe-inspiring technologies is expected to skyrocket, perhaps doubling every 5-6 months, as some staggering estimates suggest.

Faced with such rapidly increasing energy demands, the quest for more energy-efficient hardware solutions, such as photonic neural networks, has become a critical priority for researchers worldwide. A crucial step towards effectively integrating photonic circuits into neural network applications is the development of a photonic implementation for the so-called backpropagation, the gold standard in neural network training methods.

In a monumental leap forward, Sunil Pai and colleagues have designed and brought to life a cutting-edge hybrid photonic neural network (PNN) chip, capable of performing lightning-fast and efficient on-chip backpropagation training. Employing their state-of-the-art, multilayer photonic integrated circuit, Pai et al. conducted in situ backpropagation training by sending light-encoded errors backwards through the photonic neural network and measuring the optical interference with the original forward-going “inference” signal.

In a series of meticulously designed proof-of-principle experiments, the authors discovered that the PNN performed on par with digital neural network platforms, heralding a new era for scalable, energy-efficient on-chip machine learning. Charles Roques-Carmes, in a thought-provoking and insightful perspective, writes, “Photonic networks are now becoming competitive with state-of-the-art digital platforms, in terms of speed and energy efficiency.”

With this groundbreaking study, it is now anticipated that, in just a few years, large-scale hybrid and all-optical photonic chips will challenge their electronic counterparts in the realm of inference and learning of real-world AI tasks. This pioneering research has the potential to transform the landscape of artificial intelligence, opening up new possibilities for more sustainable and efficient AI technologies that will shape the future of our world.

Original Article: Backpropagation training achieved in photonic neural network

More from: Polytechnic University of Milan