<p>Reservoir computing (RC) architectures only need to train connection weights of the output layer, processing time series signals efficiently with low training cost. The proposal of delay-feedback RC provides an efficient solution for hardware implementation. However, for the processing of multidimensional time series signals, RC systems are mostly based on parallel processing schemes, which inevitably increase computational resources and complicate hardware implementation. To address these challenges, this paper proposes an efficient RC system based on two-dimensional mask processing and dynamic memristor. The input layer employs two-dimensional mask processing to convert multidimensional time series signals into one-dimensional inputs. Dynamic memristor serves as nonlinear node in the reservoir layer. In the output layer, the recursive least squares online training method is used to train the connection weights. Compared with the offline training methods, the online training method reduces the training data requirement and accelerates the output layer weight convergence significantly. With only a single dynamic memristor, the proposed RC system achieves a normalized root mean square error of <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11071_2025_11176_Article_IEq1.gif" Format="GIF" Height="17" Rendition="HTML" Resolution="72" Type="Linedraw" Width="82" /> </InlineMediaObject> <EquationSource Format="TEX">$1.21\times 10^{-7}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1.21</mn> <mo>×</mo> <msup> <mn>10</mn> <mrow> <mo>-</mo> <mn>7</mn> </mrow> </msup> </mrow> </math></EquationSource> </InlineEquation> for the short-term prediction of Hénon map, which is two orders of magnitude lower than that of the parallel reservoir computing system. For long-term prediction of Lorenz63 system, the proposed RC system demonstrates a significantly extended effective prediction time compared to other RC systems. Additionally, this system achieves <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11071_2025_11176_Article_IEq2.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="51" /> </InlineMediaObject> <EquationSource Format="TEX">$99.62\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>99.62</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> accuracy for complex time series signals classification task. Besides, the non-ideality of memristor is also considered to evaluate the performance of the proposed RC system.</p>

An efficient reservoir computing system based on 2D mask processing and dynamic memristor

期刊论文
发表日期：

2025年4月19日

查看全文

Zhuosheng Lin,
Jingliang Deng,
Zaofeng Chen,
Jilong Zhang,
Shufen Liang,
Yue Feng

摘要

Reservoir computing (RC) architectures only need to train connection weights of the output layer, processing time series signals efficiently with low training cost. The proposal of delay-feedback RC provides an efficient solution for hardware implementation. However, for the processing of multidimensional time series signals, RC systems are mostly based on parallel processing schemes, which inevitably increase computational resources and complicate hardware implementation. To address these challenges, this paper proposes an efficient RC system based on two-dimensional mask processing and dynamic memristor. The input layer employs two-dimensional mask processing to convert multidimensional time series signals into one-dimensional inputs. Dynamic memristor serves as nonlinear node in the reservoir layer. In the output layer, the recursive least squares online training method is used to train the connection weights. Compared with the offline training methods, the online training method reduces the training data requirement and accelerates the output layer weight convergence significantly. With only a single dynamic memristor, the proposed RC system achieves a normalized root mean square error of $1.21\times 10^{-7}$ $1.21 \times 10^{- 7}$ for the short-term prediction of Hénon map, which is two orders of magnitude lower than that of the parallel reservoir computing system. For long-term prediction of Lorenz63 system, the proposed RC system demonstrates a significantly extended effective prediction time compared to other RC systems. Additionally, this system achieves $99.62\%$ $99.62 %$ accuracy for complex time series signals classification task. Besides, the non-ideality of memristor is also considered to evaluate the performance of the proposed RC system.