With more connected devices, more bandwidth is increasing for tasks like telewacking and cloud computing, making it extremely difficult to manage a finite amount of wireless spectrum that all users can share.
Engineers employ artificial intelligence to dynamically manage available wireless spectrum, turning their eyes to increased latency and performance. However, most AI methods for classifying and processing wireless signals become powered and cannot work in real time.
Currently, MIT researchers are developing new AI hardware accelerators specifically designed for wireless signal processing. Optical processors perform machine learning calculations at the speed of light and classify wireless signals in nanoseconds.
The Photonic chip is about 100 times faster than the best digital alternatives, converging to around 95% accuracy in signal classification. The new hardware accelerators are also scalable and flexible, making them suitable for 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 will be particularly useful in future 6G wireless applications, such as cognitive radio, which optimizes data rates by adapting wireless modulation formats to varying wireless environments.
By enabling edge devices to perform deep learning calculations in real time, this new hardware accelerator can provide dramatic speedups in many applications beyond signal processing. For example, it can help self-driving cars respond in a flash to change in the environment, or allow smart pacemakers to continuously monitor patients' heart health.
“There are many applications that are enabled by edge devices that can analyze wireless signals. What we present in our paper can open up many possibilities for reliable AI inference in real time. This work is the beginning of something very influential (RLE), and senior author of the paper.
He was joined to the paper by his lead author, Ronald Davis III Ph.D. 24. Zaijun Chen, former MIT postdoc, is currently an assistant professor at the University of Southern California. Ryan Hamerly, visiting scientist at RLE and senior scientist at NTT research. The research is published today Advances in science.
Light speed processing
The cutting-edge digital AI accelerator for wireless signal processing transforms signals into images and runs and classifies them via deep learning models. This approach is very accurate, but the computationally intensive nature of deep neural networks makes it impossible to do with many time-sensitive applications.
Optical systems can accelerate deep neural networks by using light to encode and process data. This is less energy intensive than digital computing. However, researchers struggle to maximize the performance of general-purpose visual neural networks when used for signal processing, while ensuring that optical devices are scalable.
By developing an optical neural network architecture specifically for signal processing called multiple analog-frequency conversion optical neural networks (Maft-on), researchers have tackled the issue head-on.
The muffing results address scalability issues by encoding all signal data and performing all machine learning operations within what is called the frequency domain before the wireless signal is digitized.
The researchers designed optical neural networks to perform all linear and nonlinear operations inwards. Deep learning requires both types of manipulation.
Thanks to this innovative design, only one mufft result device per layer is required for an entire optical neural network, in contrast to other methods that require one device for an individual computational unit or “neuron.”
“You can fit 10,000 neurons into a single device and calculate the required multiplication in a single shot,” says Davis.
Researchers accomplish this using a technique called photoelectric multiplication, dramatically increasing efficiency. You can also create optical neural networks that can be easily scaled with additional layers without the need for extra overhead.
Results in nanoseconds
Maft-onl takes wireless signals as input, processes signal data, and passes information for later operations performed by edge devices. For example, by classifying signal modulation, Maft-onl allows devices to automatically guess the type of signal and extract the data they carry.
One of the biggest challenges researchers faced when designing Maft-Onn was determining how machine learning calculations could be mapped to optical hardware.
“We couldn't use the regular machine learning framework off the shelf. We had to customize it to fit our hardware and get a sense of how to misuse physics and perform the calculations we needed,” Davis says.
When testing the architecture for signal classification in simulations, optical neural networks achieve 85% accuracy in one shot and can quickly converge to 99% or more accuracy using multiple measurements. Maft-onl only required about 120 nanoseconds to run the entire process.
“The longer the measurements, the higher the accuracy you get. As the mufft calculates inferences in nanoseconds, you don't lose much speed to get more accuracy,” adds Davis.
State-of-the-art digital radio frequency devices can perform machine learning inferences in microseconds, but optics can do that in nanoseconds or even picoseconds.
Going forward, researchers want to employ what is called multiplexing schemes to allow more calculations to be performed, muffled and reduced. We also want to extend our work to a more complex, deep learning architecture that can run a trans model or LLM.
This work was funded in part by the U.S. Army Research Institute, the U.S. Air Force, the MIT Lincoln Institute, Nippon Telegraph and Telephone, and the National Science Foundation.
