Ultrafast, highly energy-efficient AI computations may become a reality with the development of a photonic processor.
Firing up the future of machine learning, scientists from MIT and colleagues have cooked up a game-changing photonic chip that's making deep neural network computations look like a walk in the park. This high-powered optical device is not only speedy but also energy-efficient— busting the limitations of traditional electronic hardware.
Most advanced machine learning applications rely on demanding deep neural network models that are pushing the boundaries of our conventional electronic computing gear. Enter photonic hardware, which performs computations using light, offering a speedy and energy-saving alternative. However, these photonic devices can't handle all neural network computations, necessitating off-chip electronics or other tricks that slow things down.
On the tantalizing promise of a decade-long research odyssey, this friendly bunch of tech whizzes has engineered a breakthrough photonic processor that boasts the ability to perform all the critical computations of a deep neural network on the chip without leaving the optical domain—until the end to read the answer, that is.
This amazing optical device managed to complete the essential computations for a machine-learning classification task in less than half a nanosecond, achieving an astounding 92% accuracy—matching the performance of traditional hardware. The chip, which is made up of interconnected modules forming an optical neural network, is fabricated using commercial foundry processes, potentially enabling scalability and integration with electronics.
The race for faster and energy-efficient AI is on, with the photonic processor offering a huge advantage for computationally demanding applications such as lidar, scientific research in astronomy and particle physics, or high-speed telecommunications.
"We can now start thinking at a higher level about applications and algorithms," raves Saumil Bandyopadhyay '17, MEng '18, PhD '23, a visiting scientist in the Quantum Photonics and AI Group within the Research Laboratory of Electronics (RLE) and a postdoc at NTT Research, Inc. Lead author of the study, Bandyopadhyay is joined by a stellar team of researchers from the group of Marin Soljačić, the Cecil and Ida Green Professor of Physics.
Deep neural networks consist of many layers of interconnected nodes (or neurons) that operate on input data to produce an output. Key operations include linear algebra matrix multiplications, transforming data as it moves from layer to layer. But deep neural networks also perform intricate non-linear operations, like activation functions, that help the model learn complex patterns and solve more challenging problems.
While their 2017 research demonstrated an optical neural network on a single photonic chip that could perform matrix multiplication with light, this device couldn’t carry out non-linear operations on the chip. Optical data needed to be converted to electrical signals and sent to a digital processor to execute non-linear operations—a power-hungry process that proved tricky to scale.
To tackle this hurdle, the team designed devices called nonlinear optical function units (NOFUs), integrating electronics and optics to implement non-linear operations on the chip. They created an optical deep neural network on a photonic chip using three layers of devices performing linear and non-linear operations.
As the system encodes the parameters of a deep neural network into light, an array of programmable beamsplitters performs matrix multiplication on the inputs. The data then pass to programmable NOFUs, which implement non-linear functions by siphoning off a bit of light to photodiodes that convert optical signals into electric current. Feeding on this method, the system achieves ultra-low latency and carries out in-situ training more efficiently with minimal energy consumption.
"This work demonstrates that computing—at its essence, the mapping of inputs to outputs—can be compiled onto new architectures of linear and non-linear physics that enable a fundamentally different scaling law of computation versus effort needed," concludes Dirk Englund, a professor in the Department of Electrical Engineering and Computer Science, principal investigator of the Quantum Photonics and Artificial Intelligence Group and of RLE, and senior author of the study.
Scaling up this device, integrating it with real-world electronics such as cameras or telecommunications systems, and exploring algorithms taking advantage of the strengths of optics to train systems faster and more energy-efficiently will be the team's focus in future work. Welcome to the photonic revolution!
- The study, led by Saumil Bandyopadhyay, has engineered a groundbreaking photonic processor that can perform all necessary computations for a deep neural network on the chip without leaving the optical domain.
- This photonic device completed a machine-learning classification task in less than half a nanosecond, achieving an impressive 92% accuracy, matching the performance of traditional hardware.
- With this photonic processor, the race for faster and energy-efficient AI is on, particularly for demanding applications such as lidar, astronomy, particle physics, and high-speed telecommunications.
- The team implemented non-linear operations on the chip using nonlinear optical function units (NOFUs), merging electronics and optics for a more energy-efficient process.
- An array of programmable beamsplitters performs matrix multiplication on the inputs, while programmable NOFUs implement non-linear functions by converting optical signals into electric current.
- Dirk Englund, a professor in Electrical Engineering and Computer Science, concludes that this work challenges the traditional scaling law of computation versus effort, suggesting a potential for photonic technology in quantum computing, artificial intelligence, and other fields.
- The researchers plan to scale up the device, integrate it with real-world electronics, and develop algorithms that take advantage of the unique properties of optics, aiming to train systems faster and more energy-efficiently—ushering in the photonic revolution.
