Development of 3D-Integrated Artificial Neural Network for Learning and Forecasting Time-Series Information
3D-integrated multilayered physical reservoir system developed for learning and forecasting time-series information
The results from Professor Wang Gun-uk and Professor J. Joshua Yang’s groups were published in Nature Communications.

▲ Choi Sang-hyeon (first author, KU-KIST Graduate School of Converging Science and Technology, University of Southern California (USC) & University of California, Santa Barbara (UCSB)),  

J. Joshua Yang (corresponding author, USC), and Wang Gun-uk (corresponding author, KU-KIST Graduate School of Converging Science and Technology/Department of Integrative Energy Engineering).

Professor Wang Gun-uk’s group at the KU-KIST Graduate School of Converging Science and Technology/Department of Integrative Energy Engineering and Professor J. Joshua Yang’s group at the University of Southern California (USC), US, implemented a vertically multilayered 3D physical reservoir array, thus presenting a novel hardware platform capable of learning and forecasting time-series information.

The research results were published online on March 6 in Nature Communications (IF=16.6), a renowned journal in multidisciplinary research.
- Title of article: 3D-integrated multilayered physical reservoir array for learning and forecasting time-series information
- Authors: Choi Sang-hyeon (first author, KU-KIST Graduate School of Converging Science and Technology, University of Southern California (USC) & University of California, Santa Barbara (UCSB)), J. Joshua Yang (corresponding author, USC), and Wang Gun-uk (corresponding author, KU-KIST Graduate School of Converging Science and Technology/Department of Integrative Energy Engineering).

Recently, time-series data, which are continuous observations collected during a specific time frame, have gradually gained more importance as a dynamic source of information in applications such as biometric analysis, weather forecasting, stock chart estimation, and hydrological inflow forecasting.

In particular, wide reservoir computing, which has been theoretically proposed, is an advanced reservoir structure that allows for processing time-series data that is difficult to learn and forecast, using multiple reservoir layers but, to date, the implementation of wide reservoir computing has been a very difficult task due not only to the absence of a high-performance physical reservoir but also to the complexity and unstable operation associated with vertically stacking multiple physical reservoir layers.

In particular, wide reservoir computing, which has been theoretically proposed, is an advanced reservoir structure that allows for processing time-series data that is difficult to learn and forecast, using multiple reservoir layers but, to date, the implementation of wide reservoir computing has been a very difficult task due not only to the absence of a high-performance physical reservoir but also to the complexity and unstable operation associated with vertically stacking multiple physical reservoir layers.

In this study, the researchers designed and manufactured a three-dimensional multilayered physical system using a next-generation electronic device called a memristor and implemented a wide reservoir computing system composed of multiple reservoir layers in the hardware (Figure 1), producing a hardware platform that can efficiently process multiple streams of time-series information with dynamicity, which is more complicated and difficult to forecast.

In particular, because the newly developed system allows for the effective understanding and learning of the subtle data features present in complex time-series data, it is highly efficient in processing complex time-series data compared to the 2D approaches that were extensively studied in the past (Figure 2). Based on this, their results are expected to play an important role in providing groundbreaking directions for physical reservoir computing that allows for the efficient processing of multiple streams of dynamic time-series information.

This study was supported by the National Research Foundation of Korea and the Korea Institute of Science and Technology (KIST) Institutional Program.

<Figure 1>

Figure 1. Representative schematic of the multilayered reservoir structure and an image of the implemented 3D multilayered physical system.

<Figure 2>

Figure 2. (Top) Results of cell position classification and prediction. Compared to existing reservoir computing based on a 2D single reservoir layer, wide reservoir computing based on the 3D multilayered physical system predicted the cell position with a high accuracy, high speed, and lower energy consumption. (Bottom) The results confirmed that the Lorenz attractor equations, which are conventionally difficult to predict, were predicted with a lower error rate using the 3D multilayered wide reservoir computing method when compared to existing single reservoir computing.