Simple Artificial Neuron Using an Ovonic Threshold Switch Featuring Spike-Frequency Adaptation and Chaotic Activity

Milim Lee, Seong Won Cho, Seon Jeong Kim, Joon Young Kwak, Hyunsu Ju, Yeonjin Yi, Byung-ki Cheong, and Suyoun Lee

Phys. Rev. Applied 13, 064056 – Published 23 June 2020

Abstract

As an essential building block for developing a large-scale brain-inspired computing system, we propose a highly scalable and energy-efficient artificial neuron device composed of an ovonic threshold switch (OTS) and a few passive electrical components. It is found that the proposed neuron device shows not only the basic integrate-and-fire function and the rate-coding property, but also the spike-frequency-adaptation (SFA) property and the chaotic activity of biological neurons, the most common features found in mammalian cortex, but they have been hard to achieve up to now. In addition, it is shown that the energy consumption of the OTS-based neuron device scales with the size of the OTS device, extrapolating both the size and the energy efficiency to the level of a biological neuron in a human brain with state-of-the-art technology. Finally, using the OTS-based neuron device combined with the reservoir computing technique, the spoken-digit recognition task has been performed with a considerable degree of recognition accuracy (94%). These results demonstrate that our OTS-based artificial neuron device is promising for the application in the development of a large-scale brain-inspired computing system.

Received 18 March 2020
Revised 8 May 2020
Accepted 26 May 2020

DOI:https://doi.org/10.1103/PhysRevApplied.13.064056

Physics Subject Headings (PhySH)

Devices

Amorphous semiconductors Artificial neural networks Chaotic systems Semiconductor compounds Spiking neurons

Interdisciplinary PhysicsCondensed Matter, Materials & Applied Physics

Authors & Affiliations

Milim Lee^1,2, Seong Won Cho^1,3, Seon Jeong Kim^1,3, Joon Young Kwak¹, Hyunsu Ju⁴, Yeonjin Yi², Byung-ki Cheong¹, and Suyoun Lee ^1,3,*

¹Center for Electronic Materials, Korea Institute of Science and Technology, Seoul 02792, Korea
²Institute of Physics and Applied Physics, Yonsei University, Seoul 03722, Korea
³Division of Nano & Information Technology, Korea University of Science and Technology, Daejon 34316, Korea
⁴Center for Opto-Electronic Materials and Devices, Korea Institute of Science and Technology, Seoul 02792, Korea

^*slee_eels@kist.re.kr

Article Text (Subscription Required)

Click to Expand

Supplemental Material (Subscription Required)

Click to Expand

References (Subscription Required)

Click to Expand

Issue

Vol. 13, Iss. 6 — June 2020

Subject Areas

Reuse & Permissions

Access Options

Author publication services for translation and copyediting assistance advertisement

Images

Figure 1
Artificial neuron based on OTS (a) An artificial neuron circuit composed of two resistors (R₁ and R₂), one capacitor (C), and one OTS device. V_mem and I_s represent the membrane potential and the spike current, respectively. (Inset) Characteristic I-V curve of an OTS device. (b) Integrate-and-fire behavior: V_mem(t) (upper panel) and I_S(t) (lower panel) of a device as a response to a dc voltage input (V_in= 5.4 V). (c) Optical microscope image of an OTS device, which has a crosspoint sandwich structure (scale bar = 20 µm). (d) Pulsed current-voltage (I-V) characteristic curves of ten OTS devices.
Reuse & Permissions
Figure 2
Rate coding (a) I_S(t) as a response to various voltage inputs (5.4–7 V, 0.2 V step). (b) Spectral density of I_S (S_Is) for the input voltages the same as (a). Curves in (a) and (b) are shifted vertically for clarity.
Reuse & Permissions
Figure 3
Spike-frequency adaptation (a) V_mem(t) and I_S(t) for V_in= 6.2 V in Fig. 2. (b) Magnified view of (a) in the range from 50 to 59 µs, where (d)–(i) correspond to the following figures (d)–(i), which explains a model suggested in this work (see the main text). ISI represents the interspike interval between adjacent spikes. (c) Firing rate (f_spike) as a function of time. The red solid line is a fitting curve modeled by the simple exponential decay f_spike(t) = f_offset + f₁ exp(−t/t₀). (d)–(i) A sequential change in V_th due to the change in the filling of trap states after switching, where the shaded region represents the filled trap states. Black dashed lines represent the virtual boundary of the off layer or the on layer, which is imagined to move along the red dotted arrows whose length represents the speed.
Reuse & Permissions
Figure 4
Chaotic activity. (a) Spectral density of V_mem (S_V_mem) of the artificial neuron obtained by fast Fourier transform, where C = 2 pF, R₁= 5 kΩ, R₂= 100 Ω. (b) Time-delay embedding plot (trajectory of [I_S(t), I_S(t + Δt), I_S(t + 2Δt)]) (Δt = 32 nsec) showing a strange attractor characterizing a chaotic system. (c) Trajectory of [I_S(t), V_mem(t), dV_mem(t)/dt] showing a similar trajectory of a biological neuron. (d) A recurrence plot of V_mem(t), where the value at time indices (i, j) is unity only if the difference between V_mem(i) and V_mem(j) is smaller than a threshold or zero otherwise, showing that the three relaxed conditions of Devaney’s chaos are shown to be satisfied (black and white dots represent unity and zero, respectively).
Reuse & Permissions
Figure 5
Energy consumption (a) V_mem (top), I_S (middle), power (bottom) of an OTS-based neuron device. In the bottom panel, the energy consumption per spike can be obtained from the area below the curve leading to about 1.2 nJ/spike. (b) Energy scalability of the OTS-based neuron device. (c), (d) Simulated V_mem (top), I_S (middle), power (bottom) of a model OTS-based neuron device with [R₁, C, C_OTS, R₂] = [5 kΩ, 100 pF, 100 pF, 0.5 kΩ] (c) and [5 kΩ, 1 pF, 1 pF, 2 kΩ] (d), respectively.
Reuse & Permissions
Figure 6
Spoken-digit recognition. (a) Schematic flowchart for spoken-digit recognition, which is composed of four steps: (1) obtaining frequency components according to the cochlear model, (2) preprocessing procedure through which the frequency components are mapped to virtual nodes (virtual neurons), where M_fθ is a masking matrix, (3) recording neuron response to the input P_τθ, and (4) training for calculating the optimum synaptic weight W_θi, where i is a class index (0–9, integer). As for the test in step (4), the pretrained W_θi is used for calculating Y_τi, whose average over τ’s, (〈Y_τi〉_τ gives the guessed digit for the maximum 〈Y_τi〉_τ. (b) A colormap of 〈Y_τi〉_τ as a function of the input spoken-digit (j = 0–9, integer) and the class index (i), where the j values are horizontally shifted in order to distinguish the same digits from different speakers (see the text). The best recognition accuracy of 94% is obtained.
Reuse & Permissions

Physical Review Applied