POTSDAM -- Erik Bollt, W. Jon Harrington Professor of Mathematics and Electrical and Computer Engineering at Clarkson University and his co-authors at Ohio State University recently published their research on next generation reservoir computing (NG-RC) in the journal Nature Communications. The publication was also just selected as the top featured article in their AI and Machine Learning section for 2021.

Reservoir computing is a machine learning algorithm developed in the early 2000s and used to solve extremely complex computing problems like forecasting dynamical systems that can change over time such as the weather. This relatively new type of computing mimics the way the human brain works and allows scientists to tackle some of the most complex information processing problems. Reservoir computing (RC) is a variation of deep learning where rather than the usual computationally intensive training phase required of most neural network methods, most of the internal weights are simply selected randomly. Nonetheless, the reservoir computing approach can perform on par with the standard fully trained neural networks. However, it was long an open question as to how and why a random method could work so well. In a paper [cite] early this year Bollt published a mathematical analysis explaining the RC successes. Now, in the recent Nature published work, Bollt and his team have further developed a next generation version, NG-RC, which is a way to make standard reservoir computing work much more efficiently, up to millions of times faster, but requiring significantly fewer computing resources and less data input needed.

“Nature Communications is a prestigious journal and we are especially honored to be chosen as a featured article in the AI and Machine Learning section. The work we are doing has the potential to let us tackle even more difficult computing problems, such as forecasting fluid dynamics someday soon,” Bollt said.

To learn more about the study, you can visit https://www.nature.com/collections/ceiajcdbeb and https://doi.org/10.1063/5.0024890