Artificial optic-neural synapse for colored and color-mixed pattern recognition
Seunghwan Seo, Seo-Hyeon Jo, Sungho Kim, Jaewoo Shim, Seyong Oh, Jeong-Hoon Kim, Keun Heo, Jae-Woong Choi, Changhwan Choi, Saeroonter Oh, Duygu Kuzum, H.-S. Philip Wong
Read Artificial optic-neural synapse for colored and color-mixed pattern recognition by Seunghwan Seo, Seo-Hyeon Jo, Sungho Kim, Jaewoo Shim, Seyong Oh, Jeong-Hoon Kim, Keun Heo, Jae-Woong Choi, Changhwan Choi, Saeroonter Oh, Duygu Kuzum, H.-S. Philip Wong in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6269540.txt.
Introduction
Seo and co-authors built an artificial synapse that can also sense colour. Their study appeared in Nature Communications. Most synapse devices only copy how nerve links change strength. This device adds light sensing on the same unit. It is built from layered materials, boron nitride and tungsten diselenide. Arranged in a network, it copies how human vision sorts coloured patterns.
The device updates its strength in a near straight line. It holds many stable states, each varying by less than one percent. It runs on tiny voltage spikes of 0.3 volts. Each spike uses only 66 femtojoules of energy. Tests showed correct sorting of coloured and mixed colour patterns. Sensing and learning in one unit saves the cost of split systems. The work points toward vision chips that learn as they see.
Catalog record
| CREATOR | Seunghwan Seo, Seo-Hyeon Jo, Sungho Kim, Jaewoo Shim, Seyong Oh, Jeong-Hoon Kim, Keun Heo, Jae-Woong Choi, Changhwan Choi, Saeroonter Oh, Duygu Kuzum, H.-S. Philip Wong |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Artificial optic-neural synapse for colored and color-mixed pattern recognition By Seunghwan Seo, Seo-Hyeon Jo, Sungho Kim, Jaewoo Shim, Seyong Oh, Jeong-Hoon Kim, Keun Heo, Jae-Woong Choi, Changhwan Choi, Saeroonter Oh, Duygu Kuzum, H.-S. Philip Wong Published in: Nature Communications DOI: 10.1038/s41467-018-07572-5 ABSTRACT The priority of synaptic device researches has been given to prove the device potential for the emulation of synaptic dynamics and not to functionalize further synaptic devices for more complex learning. Here, we demonstrate an optic-neural synaptic device by implementing synaptic and optical-sensing functions together on h-BN/WSe2 heterostructure. This device mimics the colored and color-mixed pattern recognition capabilities of the human vision system when arranged in an optic-neural network. Our synaptic device demonstrates a close to linear weight update trajectory while providing a large number of stable conduction states with less than 1% variation per state. The device operates with low voltage spikes of 0.3 V and consumes only 66 fJ per spike. This consequently facilitates the demonstration of accurate and energy efficient colored and color-mixed pattern recognition. The work will be an important step toward neural networks that comprise neural sensing and training functions for more complex pattern recognition. INTRODUCTION Since Mead1 performed the first trial to mimic the biological neural networks (BNNs) of the brain in the 1980s, extensive effort has been made to emulate BNNs by utilizing various synaptic devices2–13. In order to properly perform signal processing in BNNs, synaptic plasticity and its timing-dependent learning algorithm are two key computational parameters14. However, the scope of initial research on synaptic devices has been mainly focused on simply mimicking synaptic dynamics, such as long-term potentiation/depression (LTP/LTD)15, short-term plasticity16, and spike-timing-dependent plasticity (STDP), with filament-forming switching devices2,3 or phase change memory devices4 in array crossbar structures. Such synaptic devices benefit from superior integration capability and energy efficiency compared to the synaptic devices with lateral transistor structure. However, they suffer from nonlinear potentiation/depression characteristics17, small differences between conduction states18, and insufficient conductance states19,20. Organic5 and carbon nanotube6 transistors have recently shown their suitability for emulating of synaptic dynamics with better linear potentiation/depression characteristics and a larger number of usable conductance states (>500 distinct states), in addition to high energy efficiency (<10 pJ switching energy)21. Beyond this device-level emulation of synaptic dynamics, synaptic devices have also been used to build artificial neural networks (ANNs). Subsequently, pattern recognition tasks2–8 have been verified by these ANNs, where winner-take-all6 and perceptron networks3 are usually applied. However, none of the previous studies have tried to functionalize devices beyond their synaptic functions, for example, by merging them with biometric sensing elements such as vision, auditory, and olfactory sensors. Current works have only demonstrated signal processing in the cerebral cortex with either binary2,3 or grayscale MNIST (Modified National Institute of Standards and Technology) datasets6. The functional integration of synaptic devices with biometric sensing elements is expected to provide new opportunities for the implementation of neural networks that comprise neural sensing and training functions, thereby enabling power-efficient pattern recognition task for complex (e.g., color-mixed and voice-mixed) patterns. In this article, we demonstrate an optic-neural synaptic (ONS) device that features synaptic and optical-sensing functions. This ONS device was fabricated on a van der Waals (vdW) heterostructure (h-BN/WSe2)22,23, which does not have interfacial defects and thereby allows modulating a number of interfacial traps for achieving the synaptic functionalities. Through an optic-neural network (ONN) formed by these ONS devices, the colored and color-mixed pattern recognition capability of the human vision system is emulated. In particular, our synaptic device is investigated and compared with other devices reported heretofore, in terms of weight update linearity, number of usable conduction states, stability of each state, and energy efficiency (see Supplementary Table 1). RESULTS h-BN/WSe2 optic-neural synaptic device As shown in Fig. 1, beyond the typical emulation of synaptic dynamics, we have fabricated an ONS device by integrating a synaptic device with an optical-sensing device on the same h-BN/WSe2 heterostructure. The operation of the vdW synaptic device is based on the trapping or de-trapping of electrons in the weight control layer (WCL) on h-BN, which modulates the WSe2 channel conductivity (weight of the synapse). The vdW synaptic device will be discussed in detail in the following subsections (see below). Here, the h-BN/WSe2 photodetector resistance is modulated as a function of the incident wavelength (see Fig. 1a). The additional analysis on the h-BN and WSe2 flakes is described in Supplementary Figure 1. The h-BN/WSe2 interface was confirmed to be clearly formed without polymer residues24, owing to our transfer method based on adhesion energy engineering (see Supplementary Figure 2). Because more optical absorption occurs at a shorter wavelength (see Supplementary Figure 3a), a shorter wavelength of light decreases the resistance of the optical-sensing device under a constant drain bias (see Supplementary Figure 3b, c). This reduction in resistance means that more carriers are generated in WSe2 so that the density of carriers trapped in the WCL increases. This subsequently allows the adjustment of the synaptic dynamic properties of the ONS device according to light wavelength conditions. Here, the optical-sensing device was fabricated on the same h-BN/WSe2 heterostructure where the synaptic device was formed22,25. This was done in order for the optical-sensing device to have the proper series resistance, which is comparable to that of the synaptic device, and to enable changing the synaptic properties. If the optical-sensing device has too much larger or smaller resistance value compared to the synaptic device, the resistance control in optical-sensing device by adjusting illumination condition will be ineffective to change the synaptic properties of the ONS device. To identify the synaptic dynamics of the ONS device under different light wavelength conditions, we investigated the device’s synaptic plasticity, the postsynaptic current (PSC), and LTP/LTD. The optical-sensing device was exposed to red (λ = 655 nm), green (λ = 532 nm), and blue (λ = 405 nm) lights with an optical power of 6 mW cm−2, and we then confirmed the PSC characteristics after applying ±1 V of voltage pulse (Vpulse) to the synaptic device (see Fig. 1b and Supplementary Figure 4 for the measurement set-up). When exposed to a light with shorter wavelength, the magnitude of the synaptic current increased from 1.55 nA to 2.29 μA each, where the extracted conductance change (ΔG) also presented a significant increase from 0.78 nS to 0.74 μS by three orders of magnitude. As shown in Fig. 1c, the LTP/LTD curves were distributed in different conductance regions according to the light wavelength, but they maintained the curved shape related to nonlinearity. If the conductance state varies too nonlinearly, similar to a square root function, additional periphery circuits are typically required to manage the wide range of conductance state changes26,27. In addition, a nonlinear conductance response imposes difficulty in achieving a sufficient number of usable conductance states because a voltage pulse disables the induction of a sufficiently large conductance difference in the saturated region. Here, we defined effective conductance states as those in which ΔG exceeds a certain percentage of Gmax/Gmin (thresholdΔG as described in detail in Supplementary Figure 5). Based on thresholdΔG, we excluded the pulse signals that induced states with small ΔG unimportant for neuromorphic computing, thus reducing power consumption during computation. For the quantitative analysis of the LTP/LTD characteristics of our ONS device,