Sung Keun Shim1,Yoonho Jang1,Janguk Han1,Jeong Woo Jeon1,Cheol Seong Hwang1
Seoul National University1
Sung Keun Shim1,Yoonho Jang1,Janguk Han1,Jeong Woo Jeon1,Cheol Seong Hwang1
Seoul National University1
Recently, Reservoir Computing (RC), a temporal kernel-based computing method that processes input at the reservoir (a fixed recurrent network) and identifies it in the readout layer, has been studied due to its efficient but cost-effective learning in the machine learning field. Several attempts have been made to implement this reservoir functionality through a physical system using diffusive memristors that showed both nonlinearity and fading memory properties [1]. However, the time scale is fixed with the memristor's material properties in the previous studies, making it difficult to adjust the temporal properties of the kernel. In this study, nonvolatile W/HfO<sub>2</sub>/TiN (WHT) memristor and TiN/ZrO<sub>2</sub>/Al<sub>2</sub>O<sub>3</sub>/ZrO<sub>2</sub>/TiN (TZAZT) capacitor-based temporal kernel with an integrated 2memristor-1capacitor (2M1C) structure is proposed to solve the issue with fixed time scale.<br/><br/>HfO<sub>2</sub>-based memristor has been widely studied for various applications, including synaptic applications in neuromorphic hardwares. Especially, the WHT memristor in this study shows area-dependent electronic bipolar resistive switching behavior, which works as an analog switching device for the integrated temporal kernel. 2M1C kernel is composed of two serially connected memristors (M1, M2) and a capacitor (C1) which is connected in parallel with one of the memristors (M2). WHT memristor works as both a multi-level conductance device and a variable resistance that generates a broad range of RC delays along with C1. By modulating memristors and capacitors in the kernel circuits, time constants of the 2M1C temporal kernel can vary for 10<sup>6</sup> orders. This study analyzed how each element's characteristics affected the entire kernel properties.<br/><br/>The task of recognizing digit images in the MNIST database was conducted to verify the performance of the temporal kernel. 784 pixel data of MNIST images were binarized, chopped into n-bit pulse sequences, and fed to the kernel as an input signal. The kernel system can separate different pulse streams in terms of the number of pulses and intervals due to its unique delay dynamics through analog memristors. 2M1C kernel maps the processing results with two memristors (M1, M2; dual mapping), which projects the input signals into a higher-dimensional space. The high dimensionality significantly improves the kernel's separability, reducing the whole network's size needed for classification. Mackey-glass time series prediction test was applied to the 2M1C kernel for its data processing ability and showed excellent performance under the broad test conditions compared with previous studies [2,3].<br/><br/>2M1C temporal kernel shows that it can adapt to various tasks, and real-time data processing is possible under broad conditions by modulating M1, M2, and C1 values. The HfO<sub>2</sub>-based nonvolatile memristor acts as an analog memory device as well as the dynamics-maker. The kernel machine took 200 ns of time and ~25pJ of energy to process one input pulse, substantially reducing time and energy compared with previous studies [3]. Tunable RC delay and dual mapping techniques further contribute to the kernel's energy-efficient and accurate processing ability.<br/><br/>[1] Du, C. et al. Reservoir computing using dynamic memristors for temporal information processing. <i>Nat. commun.</i> 8.1 (2017): 1-10.<br/>[2] Jang, Y. H. et al. Time-varying data processing with nonvolatile memristor-based temporal kernel. <i>Nat. commun.</i> 12.1 (2021): 1-9.<br/>[3] Moon, J. et al. Temporal data classification and forecasting using a memristor-based reservoir computing system. <i>Nat. Electron.</i> 2.10 (2019): 480-487.