S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Dept. of Electrical and Computer Engineering (전기·정보공학부) Theses (Ph.D. / Sc.D._전기·정보공학부)
Gated Schottky Diode-type Synaptic Device with Linear Conductance Response for Hardware-based Artificial Neural Network
하드웨어 인공 신경망 구성을 위한 선형적인 전도도 특성을 가지는 게이티드 쇼트키 다이오드형 시냅스 모방 소자
- 공과대학 전기·컴퓨터공학부
- Issue Date
- 서울대학교 대학원
- Schottky junction; gated Schottky diode; effective Schottky barrier height; reconfigurability; linear conductance response; synaptic device; hardware-based artificial neural network
- 학위논문 (박사)-- 서울대학교 대학원 : 공과대학 전기·컴퓨터공학부, 2018. 2. 이종호.
- Recent developments in the Internet and sensor technologies enable the generation of diverse data and the creation of new information derived from the relationship between the data. The resulting information from various areas such as social science, economics and science and technology includes rules or patterns that were difficult to observe. Despite the current technological advances, processing large amounts of complex data and finding rules from the data to generate new information require more powerful computing capabilities. Neuromorphic computing or neural network, inspired by the structure of the biological brain, has emerged as attractive research area to satisfy the demand for computing power. Neural network performs computing tasks through connections between many neurons, and store information at the synapses that connecting neurons. Spike timing dependent plasticity and backward propagation are two typical algorithms for neural network. Spike timing-dependent plasticity-based neural network mimics the behavior of the brain in a more similar way, but the research remains at an early stage because there are still many unknown areas in terms of the behavior of the actual brain. On the other hand, backward propagation based neural network is based on mathematical theory and has matured to the point of commercialization. Neural network technology based on backward propagation is a software-based neural network because it is implemented by utilizing computing power of existing computer, therefore, the advantages of neural network, such as the low-power and high-density, cannot be exploited. Therefore, it is necessary to implement a hardware-based neural network that is different from existing computers that can efficiently implement backward propagation based neural network. The key to implement a hardware-based neural network is the effective design of synaptic device that can combine storage and computational capabilities. Synaptic device for back propagation based hardware-based neural networks should have characteristics such as low-power operation, small size, repeatability, reliable operation, and linear and symmetric conductance response according to the number of potentiation and depression pulses. Among these characteristics, the linear conductance response is an important parameter for determining training accuracy. A 2-terminal synapse device based on resistive memory technology has been proposed for a high-density synapse device array, but most resistive memories have a nonlinear conductance response. Conventional studies have attempted to solve the nonlinearity problem by using complex pulses to obtain a linear conductance response despite the burden of requiring external circuitry. However, the need for external circuitry in a hardware-based neural network means that all neurons have to contain additional circuitry or that all synapses are accessible to the public circuitry through routing. If all neurons contain additional circuitry, the area of the neural network increases, and if the synapse approaches the common circuit through the routing, then the parallel operation is not possible. Therefore, there is a need for new device that can meet the requirements of synaptic device without additional circuitry.
A novel gated Schottky diode-type synaptic device is proposed for backward propagation based hardware neural network with nearly linear conductance response in long term potentiation. The operation principle and the theoretical background of the device with linear conductance response is systematically analyzed and verified through device simulation. After that, the gated Schottky diode is fabricated with six masks using the unit processes of the conventional Si CMOS technology. The current of the proposed gated Schottky diode is small enough to be used as a low power synaptic device because the proposed device using reverse current of the Schottky diode. By applying voltage to the gate or the changing the amount of charge stored in the charge trap layer, the conductance can be changed in the gated Schottky diode. When potentiation pulses are applied to the gate consecutively, the conductance response of the proposed device is more linear than the conductivity response of other reported devices. In addition, the proposed device is fabricated by using the unit processes of Si CMOS technology, thus the device characteristics are reliable and repeatable. The operation principle, fabrication method, DC operation characteristic, pulse response and linear conductance response of the proposed gated Schottky diode is expected to contribute to the implement of the backward propagation-based hardware-based neural network.