Deep Learning with Light | MIT News
Ask the smart home device for the weather forecast, and it takes several seconds for the device to respond. One reason this latency occurs is because the connected device does not have enough memory or power to store and run the massive machine-learning model needed for the device to understand what the user is asking. The model is stored in a data center that may be hundreds of miles away, where the answer is calculated and sent to the device.
MIT researchers have created a new method for computing directly on these devices, which significantly reduces this latency. Their technology moves the memory-intensive steps of running a machine-learning model to a central server where the model’s components are encoded on light waves.
Waves are transmitted to a connected device using fiber optics, which enables tons of data to be sent lightning-fast through a network. The receiver then employs a simple optical device that performs rapid calculations using parts of the model carried by those light waves.
This technology improves energy efficiency by more than a hundred times as compared to other methods. It can also improve security, as user data does not need to be transferred to a central location for computation.
This method could enable a self-driving car to make decisions in real time, while using only a small percentage of the energy currently required by power-hungry computers. This allows the user to interact latency-free with their smart home device, use it for live video processing over a cellular network, or even enable high-speed image classification on a spacecraft millions of miles from Earth. could.
“Every time you want to run a neural network, you have to run the program, and how fast you can run the program depends on how fast you can pipe the program out of memory. Ours The pipe is massive – it’s analogous to sending a full feature-length movie over the Internet every millisecond. That’s how fast the data gets into our system. And it can calculate just as fast,” senior author Dirk Englund, an associate professor in the Department of Electrical Engineering and Computer Science (EECS) and member of the MIT Research Laboratory of Electronics.
Joining Englund on the paper is lead author and EECS graduate student Alexander Sluds; EECS graduate student Saumil Bandyopadhyay, research scientist Ryan Hammerley, as well as others from MIT, MIT Lincoln Laboratory and Nokia Corporation. research published today science,
lighten the load
Neural networks are machine-learning models that use connected nodes, or layers of neurons, to recognize patterns in datasets and perform tasks, such as classifying images or recognizing speech. But these models can have billions of weight parameters, which are numerical values that change when the input data is processed. These loads must be stored in memory. Also, the data transformation process involves billions of algebraic computations, which require a great deal of power to perform.
The process of fetching data (in this case a load of neural networks) from memory and moving them to the parts of the computer that do the actual computation is one of the biggest limiting factors for speed and energy efficiency, says Sludes.
“So our idea was, why don’t we take that heavy lifting – the process of bringing billions of weights out of memory – take it off the edge device and put it someplace where we have abundant access to power and memory, which gives Do we have the ability to bring those weights on quickly?” He says.
The neural network architecture they developed, Netcast, involves storing weights in a central server that is connected to a new piece of hardware called a smart transceiver. This smart transceiver, a thumb-sized chip that can receive and transmit data, uses a technology known as silicon photonics to fetch trillions of weights from memory each second.
It receives loads in the form of electrical signals and records them on light waves. Since the weight data is encoded as bits (1s and 0s), the transceiver converts them by switching lasers; A laser is turned on for 1 and turned off for 0. It combines these light waves and then periodically transmits them through a fiber optic network so that client devices do not need to query the server to receive them.
“Optics is great because there are so many ways to carry data within optics. For example, you can cast data on different colors of light, and this enables much higher data throughput and more bandwidth than electronics, Bandopadhyay explains.
trillion per second
Once the light waves reach the client device, a simple optical component known as a broadband “Mach-Xander” modulator uses them to perform super-fast, analog computations. This involves encoding the input data from the device, such as sensor information, onto the weights. It then sends each individual wavelength to a receiver that detects the light and measures the result of the calculation.
The researchers devised a way to use this modulator to perform trillions of multiplications per second, which increases the speed of computation on the device while using only a small amount of power.
“To make something faster, you need to make it more energy efficient. But there’s a tradeoff. We’ve built a system that can operate with about a milliwatt of power but still trillions per second.” That’s a gain of orders of magnitude, both in terms of speed and energy efficiency,” Slads says.
He tested this architecture by sending weights over an 86-kilometer fiber that connects his lab to the MIT Lincoln Laboratory. Netcast enabled machine-learning with high accuracy – 98.7 percent for image classification and 98.8 percent for numeral recognition – at a faster rate.
“We had to do some calibration, but I was surprised by how little work we had to do to get such high accuracy out of the box. We were able to achieve commercially relevant accuracy,” Hamerley says.
Going forward, the researchers want to iterate on the smart transceiver chip to achieve better performance. They want to shrink the receiver, which is currently the size of a shoe box, down to the size of a chip so that it can fit on a smart device like a cell phone.
Euan Allen, Royal Academy of Engineering Research Fellow at the University of Bath, says, “Using photonics and light as a platform for computing is a great deal of research with a potentially huge impact on the speed and efficiency of our information technology landscape.” It’s an exciting field.” who were not involved in this work. “The work of Slads et al. is an exciting step towards seeing real-world implementation of such devices, introducing a new and practical edge-computing scheme computed at very low (single-photon) light levels Exploring some fundamental limits.
Research is partially funded by NTT Research, the National Science Foundation, the Air Force Office of Scientific Research, the Air Force Research Laboratory, and the Army Research Office.
#Deep #Learning #Light #MIT #News