講演情報
[22p-A302-15]Physical reservoir based on nanoscale CNT/HfO2/CNT junctions
〇(PC)Adha Sukma Aji1, Yutaka Ohno1 (1.Inst. of Materials and Systems for Sustainability, Nagoya Univ.)
キーワード：
carbon nanotubes、reservoir computing
The high-dimensional dynamics of the reservoir computing systems are great at capturing and modeling the nonlinear dynamics, which leads to accurate prediction of the time-series data. [1,2] The use of nanomaterial-based network structure holds great potential for the realization of ultra-dense and high-dimensional physical reservoir. In this work, we utilize carbon nanotube (CNT) films which has a random network morphology. By stacking two CNT films with a thin HfO2 insulator, we utilize it as a physical reservoir capable of performing complex NARMA tasks.
We performed the autoregressive moving average (NARMA) tasks which were a commonly used benchmark for RC when evaluating the capability to replicate a higher-order dynamical model. [1] Figure 1(b) shows a reservoir output waveform after training in the NARMA2 task. With a small normalized mean square error (NMSE) of ~0.06, it is clear that our CNT/HfO2/CNT network device can perform the NARMA2 tasks. Figure 1(c) shows NMSE as a function of the order of NARMA task, showing the present device is also capable of doing higher-order NARMA tasks.
We performed the autoregressive moving average (NARMA) tasks which were a commonly used benchmark for RC when evaluating the capability to replicate a higher-order dynamical model. [1] Figure 1(b) shows a reservoir output waveform after training in the NARMA2 task. With a small normalized mean square error (NMSE) of ~0.06, it is clear that our CNT/HfO2/CNT network device can perform the NARMA2 tasks. Figure 1(c) shows NMSE as a function of the order of NARMA task, showing the present device is also capable of doing higher-order NARMA tasks.