Photo credit: www.sciencedaily.com

Advancements in Photonic Chips for Deep Learning

The latest developments in deep neural network models have revealed significant challenges for traditional electronic computing hardware, primarily due to their increasing size and complexity.

In response to these limitations, photonic hardware that utilizes light for machine-learning computations presents a promising solution, offering improvements in speed and energy efficiency. Nevertheless, certain neural network computations remain unfeasible for photonic devices, necessitating reliance on external electronic systems or other methods that can diminish overall performance.

In a significant breakthrough, a team of researchers from MIT and other institutions has engineered a new photonic chip that effectively addresses these obstacles. Their innovation involves a fully integrated photonic processor capable of executing all fundamental computations of a deep neural network directly on the chip using optical methods.

The optical device successfully completed critical computations for a machine-learning classification task in under half a nanosecond, while achieving an impressive accuracy rate of over 92 percent. This level of performance rivals that of traditional computing hardware.

Constructed from interconnected modules that represent an optical neural network, the chip is produced utilizing commercial foundry techniques, paving the way for potential scalability and integration with standard electronic systems.

Looking ahead, this photonic processor could enhance the efficiency and speed of deep learning applications across various fields, including lidar technology, scientific endeavors in astronomy and particle physics, and high-speed telecommunications.

“In many scenarios, the speed of obtaining results is just as crucial as the model’s accuracy. With our new end-to-end optical system capable of running neural networks at nanosecond speeds, we can begin to rethink potential applications and algorithms,” explains Saumil Bandyopadhyay, a lead author of the study on this innovative chip.

Bandyopadhyay collaborated with an array of researchers, including Alexander Sludds, Nicholas Harris, Darius Bunandar, and others, including senior author Dirk Englund, a professor at the Department of Electrical Engineering and Computer Science and a principal investigator at the Quantum Photonics and Artificial Intelligence Group. Their findings are published in Nature Photonics.

Harnessing Light for Machine Learning

Deep neural networks consists of numerous intertwined layers of nodes or neurons that process input data to generate outputs. A fundamental aspect of this process involves linear algebra, particularly matrix multiplication, which transforms data as it progresses through the layers.

However, in addition to linear operations, deep neural networks must also execute nonlinear operations, such as activation functions, which enable them to identify complex patterns and solve intricate problems.

In 2017, Englund’s team, in collaboration with Marin Soljačić’s lab, revealed an optical neural network functioning on a single photonic chip capable of performing matrix multiplication with light. Yet, the technology lacked the capability for onboard nonlinear operations, requiring conversion of optical data to electrical signals for processing.

“Creating nonlinearity in optics presents significant challenges, as photons interact subtly, making it power intensive to invoke optical nonlinearities—this complicates the construction of a scalable system,” Bandyopadhyay notes.

To surmount this challenge, the researchers designed nonlinear optical function units (NOFUs) that integrate both electronic and optical elements to achieve nonlinear operations directly on the chip.

By employing three layers of devices for linear and nonlinear computations, the team successfully constructed an optical deep neural network.

A Seamless Optical Integration

The novel system initiates by encoding a deep neural network’s parameters into light. Subsequently, an array of programmable beamsplitters, previously demonstrated, performs the required matrix multiplication on the incoming optical signals.

The processed data is then directed to programmable NOFUs that carry out nonlinear functions by routing a fraction of the light to photodiodes, converting the optical signals to electric currents. This efficient process minimizes the need for external amplification while keeping energy consumption low.

“By staying entirely within the optical domain until the final readout phase, we achieve ultra-low latency,” Bandyopadhyay adds.

This reduced latency allowed the researchers to effectively train their deep neural network on the chip, a method known as in situ training, typically characterized by high energy consumption in traditional digital hardware.

“This approach benefits applications involving real-time learning or in-domain processing of optical signals, such as navigation and telecommunications,” he explains.

The photonic system achieved an accuracy rate exceeding 96 percent in training tests and maintained over 92 percent accuracy during inference, comparable to conventional hardware, all while executing computations in under half a nanosecond.

Englund emphasizes that the research illustrates how computation—essentially the translation of inputs into outputs—can be adapted to new architectures utilizing linear and nonlinear physics, potentially revolutionizing the scaling of computational efforts.

Produced using the same processes that manufacture CMOS computer chips, the chip’s architecture allows for minimal fabrication errors and suggests that scaling the device for real-world applications, including integration with electronics like cameras or telecommunications systems, will be the focus of future research. The team also intends to investigate algorithms that could exploit the benefits of optical systems for faster and energy-efficient training.

This research was partially supported by the National Science Foundation, the Air Force Office of Scientific Research, and NTT Research.

Source
www.sciencedaily.com