EPFL Researchers Leverage Optics to Advance Efficient, ‘Supercomputer-Level’ Machine Learning – HPCwire

As datasets get larger and larger, the potential of machine learning insights from those datasets grows correspondingly immense – but bottlenecks in computing and data transfer speeds limit the attainment of those insights. Now, advances from researchers at the École polytechnique fédérale de Lausanne (EPFL) in Switzerland have moved one promising solution – fiber optics-based information processing – a step closer to viability.

“Light transmits information without any physical interference from cables. That’s the core advantage of optical technology when it comes to transferring data,” explained Demetri Psaltis, head of EPFL’s Optics Laboratory. “To take artificial intelligence as an example, many AI programs require accelerators to carry out rapid calculations using minimal power. For now, while optical technology could theoretically meet that need, it has not yet reached the applied stage – despite a half-century of research. That’s because optical computing and decision-making do not yet save either time or energy.”

Moser and Psaltis. Image courtesy of Alain Herzog/EPFL.

Psaltis, in collaboration with Christophe Moser’s Laboratory of Applied Photonic Devices (also at EPFL), set out to make progress on this rather large caveat. They developed a tool called the Scalable Optical Learning Operator (SOLO, for short). SOLO, powered by machine learning, can recognize and classify 2D images using a novel hybrid optical architecture that applies fibers for some tasks while retaining many of the benefits of traditional computing.

“The calculations are executed automatically by the propagation of light pulses inside the fiber,” explained Uğur Teğin, one of the co-authors of the paper. “This simplifies the computer’s architecture, retaining only a single neuronal layer, making it a hybrid system.”

The researchers tested the hybrid optical method using a dataset of lung X-rays from COVID-19 patients, attaining good accuracy (83 percent) compared to a traditional AI method. “Both systems classified the X-rays equally well,” Moser said. “However, our system consumed 100 times less energy.”

Further benchmarking bore out these results. “[The] proposed optical computing platform performs as powerful as its digital counterparts for different tasks,” the researchers wrote. “With better energy efficiency than previous proposals and a path to PetaOPs scalability, SOLO provides a promising path toward supercomputer-level optical computation.”

“[Hybrid systems] combine the bandwidth and speed of optical processing with the flexibility of electronic computing,” Psaltis said. “When coupled with artificial intelligence programs in robotics, microscopy and other visual computing tasks, these hybrid systems could achieve some of the transformative capabilities that were, for a long time, imagined as the sole purview of optical computers.”

About the research

The research discussed in this article was published as “Scalable optical learning operator” in the August 2021 issue of Nature Computational Science. The paper was written by Uğur Teğin, Mustafa Yıldırım, İlker Oğuz, Christophe Moser and Demetri Psaltis and can be accessed in full here.

Spread the love

Leave a Reply

Your email address will not be published.