Researchers at the Massachusetts Institute of Technology (MIT) have developed a nanophotonic technology that could remarkably increase the speed and efficiency of deep learning computations.
“Typically, a neural network is simulated on electronic computers,” explains Dr Yichen Shen, post-doc associate in the MIT Department of Physics. “Rather than simulate the neural network, we physically implement it by controlling the strength of the connections between nano-scale optical waveguides — which are our neurons — fabricated in the same platforms used to make the chips in your phone and computer.”
Mastering complex matrix multiplication
As analogy, the expert explain how in fact even an ordinary eyeglass lens carries out a complex matrix multiplication on the light waves that pass through it. He says the way light beams carry out computations in these photonic chips is far more complex but has a similar underlying principle: It uses multiple light beams directed in such a way that their waves interact with each other, producing interference patterns that can be interpreted to convey the result of the intended operation.
Solving the efficiency problem
Traditional computer architectures are not very efficient in processing the calculations needed for important neural-network tasks, which typically involve repeated multiplications of matrices that can become computationally intensive in conventional CPU or GPU chips. “By using light instead of electronics, we leverage the natural advantage of light on doing linear operations so that we can carry out matrix multiplications — the majority of deep learning computing — in a much faster way,” Shen says.
The future of optical computing
Shen agrees that his team’s breakthrough could impact the design of future light-based technologies and optical computing. “It will still take a lot more effort and time to make this system useful,” he says. “However, once the system is scaled up and fully functioning, it can find many user cases, such as data centers or security systems.” The system could even help to advance self-driving cars or drones and other applications that require extremely complex computations without using a lot of energy or time.
“We will scale up our chip to hundreds or even thousands of neurons and make it more practical for real case applications,” Shen shares about his team’s plans for moving ahead with the research endeavor.
The research is detailed in the article “Deep learning with coherent nanophotonic circuits,” published in the journal Nature Photonics.
Written by Sandra Henderson, Research Editor, Novus Light Technologies Today