Quantitative Relationship Between Data Dimensionality and Information Processing Capability Revealed via Principal Component Analysis for Non-Linear Current Waveforms With Non-Ideality Derived From Ionic Liquid-Based Physical Reservoir Device

Physical reservoir computing (PRC) is attracting considerable attention as a low-power, high-performance, edge artificial intelligence-friendly information technology. A physical reservoir device (PRD), which is key to PRC, converts input signals into higher-dimensional signals. In this study, we perform dimensionality analysis using principal component analysis (PCA) to investigate the data characteristics of output signals from an ionic liquid-based PRD (IL-PRD) from the viewpoint of its impact on the signal waveform and machine learning (ML) capability. More specifically, the data analysis algorism consisting of a low-rank approximation and data reconstruction is applied to the signal waveform derived from the electrochemical reactions in IL-PRD. Results show that nonlinear output signals, such as faradaic currents, increase the data dimension, thus improving the ML capability of PRC. When the data dimension increases from 1 to 9 owing to the faradaic currents, a short-term memory capacity, which is one of the indispensable performance indexes for PRC, improves by 15 times. In addition, we find that non-ideal electrical properties, such as cycle-to-cycle current variations, are included in the high-rank component of the data matrix obtained via PCA. Such non-ideality is essential for processing the raw data produced in real space because it includes the information of the input signal history as an intrinsic variability. Quantitative data dimensionality analysis, conducted using PCA, successfully explains the relationship between the ML capability of PRC and the output current waveform of the IL-PRD. These results provide guidelines for learning data selection in constructing optimal ML models.