1005 Alcyon Dr Bellmawr NJ 08031
+1 484-446-2566
support@bmcpros.com
As more and more connected devices require increasing bandwidth for tasks such as remote operations and cloud computing, managing limited wireless spectrum that can be shared by all users will become very challenging.
Engineers are using artificial intelligence to dynamically manage available wireless spectrum with an eye on reducing latency and improving performance. However, most AI methods used to classify and process wireless signals are eager and cannot run in real time.
Now, researchers at MIT have developed a new AI hardware accelerator designed for wireless signal processing. Their optical processors perform machine learning calculations at the speed of light and classify wireless signals in nanoseconds.
Photonic chips are 100 times faster than the best digital alternative, while signal classification is about 95%. The new hardware accelerator is also scalable and flexible, so it can be used in a variety of high-performance computing applications. At the same time, it is smaller, lighter, cheaper and more energy-efficient than digital AI hardware accelerators.
This device may be particularly useful in future 6G wireless applications, such as cognitive radios that optimize data rates by adjusting wireless modulation formats into changing wireless environments.
By enabling edge devices to perform deep learning computing in real time, this new hardware accelerator can provide huge acceleration in many applications beyond signal processing. For example, it can help self-driving cars to respond to environmental changes or enable smart pacemakers to continuously monitor the health of patients’ hearts.
“There are many applications that can be enabled by edge devices that can analyze wireless signals. What we propose in this article may offer many possibilities for real-time and reliable AI inference. This work is the beginning of a potentially influential start,” said Dirk Englund, a professor of electrical engineering and computer science, a senior author on electronic engineering and computer science, research in artificial intelligence and computer science, AI and computers, quantities (RLE) and the paper.
Chief writer Ronald Davis III Phd ’24 added him to the paper; Zaijun Chen is a former MIT postdoctoral fellow and now an assistant professor at the University of Southern California; Ryan Hamerly, a visiting scientist at RLE and a senior scientist at NTT Research. The study appears today Science Advances.
Speed of light processing
The most advanced digital AI accelerator for wireless signal processing converts signals into images and runs them through deep learning models to classify them. Although this approach is very accurate, the computationally intensive nature of deep neural networks makes it infeasible for many time-sensitive applications.
Optical systems can accelerate deep neural networks by using light encoding and processing data, which is also less energy-intensive than digital calculations. However, researchers strive to maximize the performance of general-purpose optical neural networks when used for signal processing while ensuring that the optical devices are scalable.
By developing an optical neural network architecture specifically for signal processing, which they call multiplication analog frequency conversion optical network (MAFT-ONN), researchers can solve the problem directly.
MAFT-ONN solves scalability problems by encoding all signal data and performing all machine learning operations in the so-called frequency domain, then before digitizing the wireless signal.
The researchers designed their optical neural network to perform all linear and nonlinear operations online. Both types of operations are deep learning.
With this innovative design, they only need to use one MAFT-ONN device per layer for the entire optical neural network, rather than other methods that require a device for each individual computing unit or “neuron”.
“We can mount 10,000 neurons onto a single device and calculate the necessary multiplication in a single lens,” Davis said.
The researchers used a technique called photoelectric multiplication to achieve this, which significantly improved efficiency. It also allows them to create an optical neural network that can be easily scaled with other layers without additional overhead.
Cause nanoseconds
Maft-onn uses wireless signals as input, processes signal data, and then passes information to subsequent operations, and the edge device performs. For example, by classifying the modulation of the signal, MAFT-ONN will enable the device to automatically infer the type of the signal to extract the data it carries.
One of the biggest challenges researchers face when designing Maft-Onn is determining how machine learning computing is mapped to optical hardware.
“We can’t just take a normal machine learning framework out of the shelf and use it. We have to customize it to fit the hardware and figure out how to leverage physics so that it can perform the calculations we want,” Davis said.
When they tested their architecture in signal classification in simulations, the optical neural network achieved 85% accuracy in a single lens, and using multiple measurements can quickly converge to more than 99% accuracy. MAFT-ONN takes only about 120 nanoseconds to perform the entire process.
“The longer you measure, the higher the accuracy you will get. Because of the inferences that MAFT-ONN calculates nanoseconds, you don’t lose too much speed to get higher accuracy.”
While state-of-the-art digital RF devices can perform machine learning inference in microseconds, optics can be performed in nanoseconds or even picseconds.
Moving forward, researchers hope to adopt a scheme called multiplexing so that they can perform more computations and scale up MAFT-ONN. They also want to extend the work to more complex deep learning architectures that can run Transformers or LLM.
This work is partly funded by the U.S. Army Research Laboratory, the U.S. Air Force, MIT Lincoln Laboratory, Japan Telegram and Telephone, and the National Science Foundation.