Pioneering study
New Possibilities for Reservoir Computing
- von Ulrike Bohnsack
- 03.07.2024
Speech recognition, weather forecasts, smart home applications: Artificial intelligence and the Internet of Things are conquering our everyday lives. Systems based on reservoir computing are very promising. The research group led by Prof Dr Karin Everschor-Sitte, Theoretical Physics at the UDE, is conducting research in this area. It is primarily investigating new possibilities for reservoir computing, for example using magnetic materials. Now, in a pioneering study together with specialists from the field of ferroelectric materials, the team has shown that these systems are also suitable for processing complex data faster and more efficiently. The results have now been published in Nature Reviews Physics.
Put simply, reservoir computing utilises a large network (the reservoir) to convert complex tasks into a form that is easy to process. "Any system that has four core properties is suitable as a physical reservoir: Complexity, short-term memory, non-linearity and reproducibility," explains UDE professor Karin Everschor-Sitte. "Magnetic patterns on the nanoscale are very interesting, especially so-called skyrmions. These magnetic vortices can be moved and excited - for example by electrical currents, temperature, voltage or light pulses - causing them to grow, shrink or deform. Because these systems are easy to manipulate and measure, energy-efficient and easy-to-control reservoirs can be constructed. They are compatible with our current computer hardware."
With their research, Everschor-Sitte and her team laid the foundation for magnetic reservoir computing around seven years ago. Now, in cooperation with colleagues from the Norwegian University of Science and Technology (NTNU), they have developed the idea of a new variant: ferroelectric reservoir computing, which is based on the special properties of ferroelectric materials: "They can store energy well, they can change their electrical polarisation, for example through an electric field or temperature; they have fast switching behaviour and they can be mechanically deformed," says co-author Dr Atreya Majumdar (UDE), listing the advantages.
Because both magnetic and ferroelectric materials are so versatile, the systems can also be multidimensional and hybrid: This means they have much greater ability to process data, recognise complex patterns and display information. Past inputs can also be better stored and utilised, which is particularly important for time-dependent data.
With the new possibilities of reservoir computing, more powerful applications could be developed in the future that have to do with speech and image recognition, sensor technology or embedded systems. "Smart home applications and the Internet of Things need compact systems that are fast and consume little energy," says Majumdar. "Reservoir computing is a promising solution here."
Original Publication: https://www.nature.com/articles/s42254-024-00729-w
Image: Yin and yang from ferroelectric (blue) and magnetic patterns (red).
Further Information:
Prof. Dr. Karin Everschor-Sitte, Theoretical Physics, +49 203/37 9-4720, karin.everschor-sitte@uni-due.de
Dr. Atreya Majumdar, Theoretical Physics, +49 203/37 9-4717,atreya.majumdar@uni-due.de
Editor: Ulrike Bohnsack, +49 151-74448046, ulrike.bohnsack@uni-due.de