Backpropagation training achieved in photonic neural network

Neural networks made from photonic chips can be trained using on-chip backpropagation – the most widely used approach to training neural networks, according to a new study. The findings pave the way toward developing optically driven and energy-efficient machine learning technologies that reduce both the carbon footprint and costs of AI computation. Neural networks (NNs) are an approach to machine learning conceptually inspired by the biology of the brain and have become a mainstay in various modern scientific and commercial AI technologies, including the widely discussed ChatGPT architectures. However, as NNs continue to advance and become even more widespread, the energy required to power the technologies is expected to grow exponentially, perhaps doubling every 5-6 months, as some estimates suggest. Such rapidly increasing energy costs necessitate a shift toward more energy-efficient hardware solutions, such as photonic NNs. One of the first things needed to effectively adopt photonic circuits in NN applications is to develop a photonic implementation for so-called backpropagation, the most widely used NN training method. Here, Sunil Pai and colleagues describe a hybrid photonic neural network (PNN) chip that can perform fast and efficient on-chip backpropagation training. Using their multilayer photonic integrated circuit, Pai et al. performed 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 proof-of-principle experiments, the authors found that the PNN performed comparably to digital NN platforms, suggesting a route for scalable, energy-efficient on-chip machine learning. “Photonic networks are now becoming competitive with state-of-the-art digital platforms, in terms of speed and energy efficiency,” writes Charles Roques-Carmes in a related Perspective. “It is hoped that in the next few years, large-scale hybrid and all-optical photonic chips will rival their electronic counterparts in inference and learning of real-world AI tasks.”