A photonic processor can improve the processing of 6G wireless signal | Myth news

Because more connected devices require the growing bandwidth of tasks such as Torking and cloud processing, management of finished wireless spectrum available to all users will be extremely difficult.

Engineers use artificial intelligence to dynamically manage the available wireless spectrum, with a view to reducing delays and increasing performance. But most AI methods for the classification and processing of wireless signals are hungry for energy and cannot work in real time.

Now scientists MIT has developed an innovative accelerator of AI equipment, which is specially designed for wireless signal processing. Their optical processor performs calculations of machine learning at the speed of light, classifying wireless signals during nanoseconds.

The photo chip is about 100 times faster than the best digital alternative, while at the same time coincides with about 95 percent accuracy in the signal classification. The new hardware accelerator is also scalable and flexible, so you can use it for various high -performance computing applications. At the same time, it is smaller, lighter, cheaper and more energy -efficient than AI digital hardware accelerators.

The device can be particularly useful in future 6G wireless applications, such as cognitive radios, which optimize data speeds, adapting wireless modulation formats to the changing wireless environment.

By enabling the device to be edited in real time, this new hardware accelerator can ensure dramatic acceleration in many applications except for signal processing. For example, it can help autonomous vehicles to make fractional reactions to environmental changes or enable intelligent steady continuous monitoring of the patient's heart.

“There are many applications that would be enabled by EDGE devices that are able to analyze wireless signals. What we presented in our article could open many possibilities of inference in real time and reliable inference AI. This work is the beginning of something that could be quite influential,” says Dirk Englund, profile in Mit Department of Engineerical and computer science, in the field of factory, artificial and artificial scientific and artificial work and science. Electronics (RLE) and the older author paper.

He is joined by the main author of Ronald Davis III PhD '24; Zajun Chen, a former postdoc mit, which is currently an adjunct at the University of Southern California; and Ryan Hamerly, visiting a scientist in RLE and senior scientist at NTT Research. The study appears today in Scientific progress.

Speed ​​processing

The most modern AI digital accelerators for wireless signal processing convert the signal to the image and run it through a deep learning model to classify it. Although this approach is very accurate, intensively computing the nature of deep neural networks makes it impossible to many sensitive time.

Optical systems can accelerate deep neural networks by encoding and processing data using light, which is also less energy -saving than digital processing. But scientists have tried to maximize the performance of optical general purpose neural networks when it is used for signal processing, while providing a scalable optical device.

When developing the optical architecture of the neural network especially for signal processing, which they call a multiplicative analogue optical transformation of the neuronal neuronal network (MAFT-ONN), scientists solved this problem.

MAFT-ONN solves the problem of scalability by encoding all signal data and performing all machine learning operations as part of the frequency domain-wireless signals will be digitized.

Scientists have designed their optical neural network to perform all line and non -linear operations in the line. Both types of operations are required for deep learning.

Thanks to this innovative design, they only need one maft-ann device for a layer for the entire optical neural network, unlike other methods that require one device for each computing unit or “neuron”.

“We can fit 10,000 neurons on one device and calculate the necessary multiplication in one shot,” says Davis.

Scientists do this using a technique called photoelectric multiplication, which significantly increases performance. It also allows them to create an optical neural network that can be easily scaled with additional layers without the requirement of an additional cost.

Results in nanoseconds

MAFT-INN takes a wireless signal as input data, processes signal data and provides information for later operations performed by the EDGE device. For example, by classifying signal modulation, MAFT-ONN would allow the device to automatically apply with the type of signal to separate the data it transfers.

One of the biggest challenges that the scientists faced when designing the Maft-Onn was to determine how to map the calculations of machine learning to optical equipment.

“We could not just take a normal machine learning frame from the shelf and use it. We had to adapt it to the equipment and come up with how to use physics to perform the calculations we wanted,” says Davis.

When they tested their architecture to the signal classification in simulations, the optical neural network reached an 85 % accuracy in one shot, which can quickly converge with over 99 -percentage accuracy using many measurements. Maft-Onn required only about 120 nanoseconds to perform the entire process.

“The longer you measure, the higher the accuracy you will receive. Because the maft-ons calculates conclusions in nanoseconds, you do not lose high speed to get higher accuracy,” adds Davis.

While the latest digital radio devices can perform machine inference in microseconds, optics can do it in nanoseconds and even pikoseconds.

Going further, scientists want to use so-called multiplex programs so that they can perform more calculations and scale maft-on. They also want to expand their work to more complex deep learning architecture, which can run transformer models or LLM.

These works were partly financed by the American research laboratory, American air force, myth of Lincoln Laboratory, Nippon Telegraph and Telefone and the National Science Foundation.