Forming neural networks not with electricity but with active colloidal particles
In a groundbreaking study, physicists at Leipzig University have introduced a new kind of neural network, diverging from traditional microelectronic chips. This innovative network operates through active colloidal particles instead of electricity. Their research, published in the acclaimed Nature Communications, illustrates how these microparticles can be employed in artificial intelligence and time series prediction.
Professor Frank Cichos, leading the research team, explained that their neural network is part of physical reservoir computing. This field utilizes the dynamics of physical processes like water surfaces and bacterial movement for computations. Supported by ScaDS.AI and funded by the German government’s AI Strategy, this research is a significant leap in AI technology, offering a novel computation method.
The team's approach involves synthetic self-propelled microparticles, meticulously engineered to perform calculations while minimizing disruptive effects such as environmental noise. These colloidal particles, finely dispersed in mediums like liquids, form the basis of their experiments. The researchers have developed units comprising plastic and gold nanoparticles, creating a system where one particle's movement around another is laser-driven. This setup has proved beneficial for reservoir computing, as each unit processes information contributing to the overall computational system.
Dr. Xiangzun Wang from the team highlights that the system’s training is crucial for specific calculations, much like other neural networks. A significant focus of their research was managing noise, an inherent challenge due to the minuscule size of the particles in water. They discovered that utilizing past states of the reservoir can enhance performance, allowing for smaller reservoirs under noisy conditions. This finding is a substantial contribution to the field, offering a way to optimize reservoir computation and reduce noise impact.
Webdesk AI News : Neural Network formed with Active Microparticles, January 29, 2024