Reservoir computing with dielectric relaxation at an electrode–ionic liquid interface

Abstract

A physical reservoir device with tunable transient dynamics is strongly required to process time-series data with various timescales generated in the edge region. In this study, we proposed using the dielectric relaxation at an electrode–ionic liquid (IL) interface as the physical reservoir by making the most of designable physicochemical properties of ILs. The transient dynamics of a Au/IL/Au reservoir device were characterized as a function of the alkyl chain length of cations in the IL (1-alkyl-3-methylimidazolium bis(trifluoromethane sulfonyl)imide). By considering a weighted sum of exponentials expressing a superposition of Debye-type relaxations, the transient dynamics were well reconstructed. Although such complex dynamics governed by multiple relaxation processes were observed, each extracted relaxation time scales with a power law as a function of IL’s viscosity determined by the alkyl chain length of cations. This indicates that the relaxation processes are characterized by bulk properties of the ILs that obey the widely received Vogel-Fulcher-Tammann law. We demonstrated that the 4-bit time-series signals were transformed into the 16 classifiable data, and the data transformation, which enables to achieve higher accuracy in an image classification task, can be easily optimized according to the features of the input signals by controlling the IL’s viscosity.

Edge-of-chaos learning achieved by ion-electron–coupled dynamics in an ion-gating reservoir

“…While the performance was not as good as that achieved by a typical three-layer NN, the size of the network in the present study (1960) is far smaller than in a three-layer NN (784,000). Compared to the recognition accuracies of other physical reservoirs (83 to 90.2%) (13,16,18,43), that of IGR is similar or slightly better. However, while the advantage of IGR is minor for such a relatively easy task, IGR showed very good computational performance on more difficult time series data analysis tasks that require superior properties, such as reservoir diversity, which is discussed below.…”

“…The size of the IGR and NNs are given in parentheses. The recognition accuracies of other physical reservoirs, such as memristors [magnetic skyrmion memristor (MSM)] ( 13 ), WO x ( 16 ), SiO x -Ag ( 18 ), and ionic liquid (IL) ( 43 ), are shown for comparison.…”

“…The IGR can map time series data in high-dimension feature space by using I D response, with asymmetric relaxation and spikes, the characteristics of which are widely modified by the input history. To investigate a function of IGR as a physical reservoir, we performed a handwritten-digit recognition task ( 13 , 16 , 18 , 43 ). Figure 1B is an example of the 28 × 28 pixel input digit “6” from the Modified National Institute of Standards and Technology (MNIST) database ( 44 ).…”

“…These values were used as the reservoir output to train and test the readout network. Similar methods have been used elsewhere ( 13 , 16 , 18 , 43 ). The details of the procedure used are given in Materials and Methods.…”

Reservoir computing with the electrochemical formation and reduction of gold oxide in aqueous solutions with a three-electrode electrochemical setup

“…We conducted an image classication task using the current transient observed according to a previously reported procedure. 10,20,21 The details are provided in the ESI. † In brief, images of handwritten numbers taken from the MNIST database were converted into a one-dimensional binary array, that is, a pulse train.…”