Artificial intelligence powered by neural networks conducts digital calculations using microelectronic chips. Physicists at Leipzig University have developed a unique neural network that operates with active colloidal particles instead of electricity. This innovative approach, detailed in Nature Communications, demonstrates the potential of microparticles as a physical system for artificial intelligence and time series prediction.
“Our implementation involves synthetic self-propelled microparticles, just a few micrometers in size,” explains Cichos. “We’ve proven that these particles can perform calculations while introducing a method to mitigate disruptive effects, such as noise, in their movement.” Colloidal particles are finely dispersed in solid, gas, or liquid mediums.
For their experiments, the physicists designed minuscule units composed of plastic and gold nanoparticles, where one particle orbits another under laser propulsion. These units possess specific physical properties that make them ideal for reservoir computing.
“Each unit is capable of processing information, and when combined, they form a ‘reservoir.’ By manipulating the rotational motion of the particles in the reservoir with an input signal, we obtain the result of a calculation,” explains Dr. Xiangzun Wang. “Similar to traditional neural networks, this system requires training for specific tasks.”
The researchers focused on addressing noise issues. “Due to the extremely small particles in water, our system is susceptible to significant noise, akin to the noise experienced by all molecules in the brain,” says Cichos.
2024-01-29 11:41:03
Link from phys.org
