Amorphous InGaZnO-Based Synaptic Device for Artificial Neural Network Application = 비정질 인듐-갈륨-아연-산화물 기반의 시냅스 소자를 이용한 인공 신경망 응용

다국어 초록 (Multilingual Abstract)

In this paper, a memristor, an oxide InGaZnO (IGZO) analog memory for constructing an artificial neural network (ANN) for neuromorphic computing, was fabricated and analyzed. In addition, three types of IGZO memristors were manufactured, each implementing an in-memory computing application.
First, instead of applying non-identical pulses such as incremental step pulse programming (ISPP) that cause hardware burden, a synaptic IGZO transistor and an IGZO memristor were combined to improve the symmetry and linearity characteristics of the synaptic weight update characteristics. As a result, edge devices using 1T-1M block and on-chip learning have improved MNIST pattern classification accuracy using deep neural networks (DNN).
Second, the 10 x10 crossbar array structure IGZO memristor was processed on a flexible substrate for wearable healthcare and IoT applications. In addition, it has been confirmed that 2x2 and 4x4 pattern classification by binary neural network (BNN) neuromorphic computing is well implemented even under mechanical bending stress.
As the last application, an amorphous IGZO-based memristor that responds sensitively to light, especially blue light, has been developed. For circadian rhythm diagnosis, current and conductivity modulation of memristor synaptic device changes according to light, particularly blue light, was confirmed. In addition, a leaky integration & fire (LIF) neuron spiking neural network (SNN) circuit that discriminates color temperature and illuminance using current changes according to light were implemented by combining an IGZO memristor serves as the optical synapse and a HfO2 conducting bridge random access memory (CBRAM) based threshold switch device serves as a neuron. It is found that the circadian light meter (CLM) discriminates the correlated color temperature (CCT) at a fixed visual illuminance and predicts the influence of light pollution. Furthermore, oxide semiconductor synapse/neuron CLM enables the integration with CMOS back end of line (BEOL) and is compatible with spiking neural network (SNN) signal processing. Therefore, it is potentially useful for edge-computing for circadian rhythm diagnosis.

국문 초록 (Abstract)

본 논문에서는 뉴로모픽 컴퓨팅을 위한 Artificial Neural Network (ANN) 를 구성을 위한 산화물 InGaZnO 아날로그 메모리인 memristor를 제작 및 분석했다. 또한 3 가지의 InGaZnO memristor를 제작하여 각각, In-memory computing application을 구현하였다. Hardware burden을 유발하는 non-identical pulses를 인가하는 대신, 시냅틱 IGZO transistor와 IGZO memristor를 결합하여 synaptic weigth update 특성의 대칭성 및 선형성 특성을 개선하였다. 이를 통해, Deep Neural Network (DNN)을 이용한 on-chip learning MNIST 패턴 정확도를 증가시켰다. 또한 웨어러블 헬스 케어 장비 및 IOT 응용을 위한 유연 기판 위에 memristor를 제작하였고, vector-matrix multiplication (VMM)을 이용한 Binary Neural Network (BNN) 뉴로모픽 컴퓨팅이 bending과 같은 기계적 스트레스에도 잘 구현됨을 확인하였다. 마지막으로 일주기리듬 진단을 위해 memristor synapse 소자의 빛에 따른 전류 변화 및 conductance modulation을 확인하였다. 또한 memristor 소자와 CBRAM 소자를 결합하여 빛에 따른 전류 변화를 이용하여 색온도 및 조도를 분간하는 Spiking Neural Network (SNN) 회로를 구현하였다. 이를 통해서 엣지 컴퓨팅을 이용하여 써카디안 리듬 진단의 가능성을 보였다.

목차 (Table of Contents)

• CHAPTER 1. Introduction 1
• CHAPTER 2. The fabrication process of synapse device 3
• CHAPTER 3. Improvement of the symmetry and linearity of synaptic weight update 5
• 3.1 Amorphous InGaZnO memristor and synaptic transistor 5
• 3.2 Combining the InGaZnO synaptic transistor and memristor 9
• CHAPTER 4. Neuromorphic Computing Using Vector-Matrix Multiplication (VMM) 13
• 4.1 Electrical characteristic of flexible memristor 13
• 4.2 Vector-matrix multiplication for binary neural network 15
• CHAPTER 5. Leaky integrate and fire synapse-neuron circuit for circadaian rhythm diagnosis 18
• 5.1 Motivation of circadian rhythm diagnosis 18
• 5.2 Photo-sensitive InGaZnO memristor 21
• 5.3 Synapse-neuron circuit for circadian rhythm diagnosis 28
• CHAPTER 6. Conclusion 33
• Reference 34
• List of Published Journal Papers 40
• List of Published Conference Papers 42
• Abstract (Korean) 43