As the more connected devices require an increasing amount of bandwidth for tasks such as telework and cloud computing, it will become extremely difficult to manage the limited amount of wireless spectrum available for all users.
Engineers use artificial intelligence to dynamically manage the available wireless spectrum, with a view to reducing latency and stimulating performance. But most AI methods to classify and treat wireless signals are swallowed with power and cannot operate in real time.
Now MIT researchers have developed a new AI hardware accelerator specially designed for wireless signal processing. Their optical processor performs automatic learning calculations at the speed of light, classifying wireless signals in a question of nanoseconds.
The photonic chip is approximately 100 times faster than the best digital alternative, while converging towards an accuracy of around 95% in the classification of signals. The new equipment accelerator is also scalable and flexible, it could therefore be used for a variety of high performance computer applications. At the same time, it is smaller, lighter, cheaper and more energy efficient than digital AI hardware accelerators.
The device could be particularly useful in future 6G wireless applications, such as cognitive radios which optimize data flow rates by adapting wireless modulation formats to the changing wireless environment.
By allowing a on-board device to carry out in-depth learning calculations in real time, this new equipment accelerator could provide spectacular accelerates in many applications beyond the processing of the signal. For example, it could help autonomous vehicles to make fractional reactions to environmental changes or to allow stimulating stimulators to continuously monitor the health of a patient's heart.
“There are many applications that would be activated by EDGE devices which are capable of analyzing wireless signals. What we have presented in our article could open many possibilities for real -time and reliable inference. This work is the beginning of something that could be quite impactful “electronic (rle), and principal author of the paper.
He is joined on the newspaper by the main author Ronald Davis III Phd '24; Zaijun Chen, a former MIT postdoc who is now a deputy professor at the University of Southern California; And Ryan Hamerly, scientist invited to Rle and principal scientist at NTT Research. Research appears today in Scientific advances.
Light treatment
The cutting-edge digital AI accelerators for wireless signal processing convert the signal into an image and run it via a deep learning model to classify it. Although this approach is very precise, the intensive nature in calculating deep neurons networks makes it unrealizable for many applications sensitive to time.
Optical systems can speed up deep neural networks by coding and processing data using light, which is also lower of energy than digital computer science. But the researchers found it difficult to maximize the performance of optical neural networks for general use when used for signal processing, while ensuring that the optical device is scalable.
By developing an optical neural network architecture specifically for the processing of the signal, which they call a network of optical neurons of multiplicative analog frequency (MAFT-SNON), the researchers resolved this problem on the front.
The MAFT -RONN tackles the problem of scalability by coding all signal data and by performing all automatic learning operations in what is known as frequency domain name – before wireless signals are digitized.
The researchers have designed their network of optical neurons to carry out all linear and non -linear operations online. The two types of operations are necessary for in -depth learning.
Thanks to this innovative conception, they only need a single Maft-Snor device per layer for the entire optical neural network, as opposed to other methods which require a device for each unit of individual calculation, or “neuron”.
“We can install 10,000 neurons on a single device and calculate the necessary multiplications in a single time,” explains Davis.
Researchers perform this using a technique called photoelectric multiplication, which considerably increases efficiency. It also allows them to create an optical neural network which can be easily widened with additional layers without requiring additional general costs.
Leads nanoseconds
MAFT-RONN takes a wireless signal in input, processes signal data and transmits the information for subsequent operations that the EDGE device performs. For example, by classifying the modulation of a signal, MAFT-SNON would allow a device to automatically deduce the type of signal to extract the data it carries.
One of the biggest challenges that researchers were faced during the design of MAFT-SNON was to determine how to map automatic learning calculations with optical equipment.
“We couldn't just take a normal automatic learning framework on the shelf and use it. We had to personalize it to adapt to the equipment and determine how to use physics so that it performs the calculations we wanted, ”explains Davis.
When they have tested their architecture on the signal classification in simulations, the optical neural network has reached 85% precision in a single time, which can quickly converge towards an accuracy of more than 99% using several measures. Maft-Ronn only requires about 120 nanoseconds to perform an entire process.
“The longer you measure, the higher you will get.” Because Maft-Ronn calculates the inferences in the nanoseconds, you do not lose much speed to gain more precision, ”adds Davis.
While advanced digital radical radiofrequencies can carry out automatic microsecond learning inference, optics can do so in nanoseconds or even picoseconds.
In the future, researchers want to use what are called multiplexing patterns so that they can perform more calculations and evolve the MAFT-INN. They also want to extend their work to more complex learning architectures that could execute models of transformers or LLM.
This work was funded, in part, by the US Army Research Laboratory, the US Air Force, the Lincoln Laboratory, Nippon Telegraph and Telephone, and the National Science Foundation.
