Research lays foundation for reconfigurable neuromorphic building blocks

Optical neural networks have the potential to provide high-speed, high-capacity solutions needed to tackle difficult computing tasks. However, further progress is needed to realize its full potential. One of the challenges is the reconfigurability of optical neural networks.

A research team from the Stiller Research Group at the Max Planck Institute for Photoscience, in collaboration with the England Research Group at the Massachusetts Institute of Technology, is developing new reconfigurable neuromorphic building blocks by adding new elements. We have succeeded in laying the foundation. Dimensions to optical machine learning: Sound waves. Their discovery is nature communications.

Researchers use light to generate temporary acoustic waves within optical fibers. Sound waves generated in this way enable recurrent functionality in optical fibers for communications, which is essential for interpreting contextual information such as language, for example.

Artificial intelligence is now commonplace and helps us get through our daily tasks. Language models such as ChatGPT can help reduce administrative overhead by creating naturally formulated text and summarizing paragraphs in a structured way. The disadvantage is that it requires a huge amount of energy. This means that as these intelligent devices evolve, new solutions will be needed to speed up signal processing and reduce energy consumption.

Neural networks have the potential to form the backbone of artificial intelligence. Building these as optical neural networks based on light rather than electrical signals allows them to process large amounts of data at high speed and with great energy efficiency. However, to date, many experimental approaches to implementing optical neural networks have relied on fixed components and stationary devices.

Now, an international research team led by Birgit Stiller at the Max Planck Institute for Photoscience, in collaboration with Dark England at the Massachusetts Institute of Technology, has discovered how to build reconfigurable building blocks based on sound waves for optical machine learning. discovered. In their experimental approach, the researchers used hair-thin optical fibers, which are already used around the world for high-speed internet connections.

Key to the invention is the optically driven generation of traveling acoustic waves that drive the subsequent computational steps of the optical neural network. Optical information is processed and associated with acoustic waves. The transmission time of sound waves is much longer than that of optical information streams. Therefore, they can remain in the optical fiber longer and be connected to each subsequent processing step in sequence. The uniqueness of this process lies in the fact that it is completely controlled by light and does not require complex structures or transducers.

“We are very excited that we have embarked on this new field of research that pioneers the use of sound waves to control optical neural networks. Our research results will lead to new building blocks for new photonic computing architectures. “It has the potential to stimulate block development,” he says. Dr. Birgit Stiller, Head of the Quantum Photoacoustics Research Group.

The first building block the team experimentally demonstrated was a recurrent operator, a widely used technology in the field of recurrent neural networks. This allows you to link a series of computational steps, thus providing context for each single computational step performed.

For example, in human languages, the order of words determines the meaning of a sentence. For example, her two sentences, “She decided to research her problem'' and “She decided to try her research,'' are made up of the same words but have different meanings. it's different. This is because word order creates different contexts.

Traditional fully connected neural networks on computers face challenges in capturing context because they require access to memory. To overcome this challenge, neural networks are equipped with iterative operations that enable internal memory and capture contextual information. While these recurrent neural networks are easy to implement digitally, similar implementations in optics have been difficult and have traditionally relied on artificial cavities to provide memory.

Researchers are currently implementing recurrent operators using sound waves. As a result, the photoacoustic recurrent operator (OREO) exploits the unique properties of optical waveguides without the need for artificial reservoirs or newly fabricated structures.

OREO has the advantage of being fully optically controlled, making the optoacoustic computer pulse-by-pulse programmable. For example, researchers used it to optically implement recurrent dropout for the first time. This is a tuning technique previously used only to improve the performance of digital recurrent neural networks. OREO has been used to distinguish up to 27 different patterns, demonstrating its ability to handle context.

“OREO's all-optical control is a powerful feature, especially the ability to program the system pulse-by-pulse, which adds several degrees of freedom.Using sound waves for optical machine learning is a game-changer. I'm very excited to see how this field evolves in the future,” says Steven Becker, a doctoral student in the Stiller Institute.

In the future, the use of sound waves in optical neural networks will unlock a new class of spontaneously reconfigurable optical neuromorphic computing, enabling large-scale in-memory computing in today's telecommunications networks. It may become. Additionally, on-chip implementations of optical neural networks can benefit from this approach, which can be implemented in photonic waveguides without additional electronic controls.

“Optical machine learning may have great potential when it comes to parallel processing and energy-efficient manipulation of information. Adding sound waves creates a fully optically controlled and easy-to-operate toolkit. You can contribute to this effort by using ,” says Dr. Birgit Stiller. .