Max Planck Institute for the Science of Light in Germany, in collaboration with researchers from the Massachusetts Institute of Technology in the United States, has laid the groundwork for reconfigurable neural morphic modules by adding a sonic dimension to photonics machine learning. This achievement is crucial for efficiently interpreting contextual semantic information in generative artificial intelligence (AI). The research findings were published on April 17th on the website of the American Association for the Advancement of Science.
Models like ChatGPT can generate natural language text and summarize paragraphs in a structured manner. However, a significant drawback is the substantial energy required to achieve this. Consequently, as these intelligent devices advance rapidly, new solutions are necessary to accelerate signal processing and reduce energy consumption.
Neural networks are considered pivotal for AI. Constructing them as optical neural networks based on light rather than electrical signals enables rapid and efficient processing of large amounts of data. However, many experimental approaches to implementing optical neural networks to date have relied on fixed components and stable devices.
The research team has devised a method for constructing reconfigurable modules based on sound waves for photonics machine learning. The key to this research lies in the traveling sound waves generated by light, which can manipulate subsequent computational steps of optical neural networks. Unlike the flow of optical information, sound waves have significantly longer transmission times, allowing them to be retained in optical fibers for a longer duration and sequentially linked to each subsequent processing step.
The team demonstrated the first constructed module — a recurrent operator, widely used in the field of recurrent neural networks. It enables linking a series of computational steps and provides context for each executed computational step.
The optoacoustic recurrent operator leverages the inherent properties of optical waveguides without the need for artificial reservoirs or new manufacturing structures. It has been used to distinguish up to 27 different patterns, demonstrating its ability to efficiently process context while conserving energy.
