뉴로모픽 응용을 위한 나노스케일 차세대 메모리 소자 기반 인공 뉴런

Abstract

Artificial intelligence (AI) is a core technology of the fourth industrial revolution with developments of the internet and sensors. The technology of AI makes revolutionary changes in a wide range of fields such as visual/auditory recognition, language translation, automatic driving, biosensor, smart robot. According to the inspired cognitive function of the human brain, various artificial neural networks (ANN) have been proposed such as deep neural network (DNN), spiking neural network (SNN), oscillator neural network (ONN), and recurrent neural network (ONN). However, it is efficient for conventional von-Neumann digital computing to perform analogous processing of ANN owing to a bottleneck in data transformation between memory and processor. Thus, bio-inspired analog computing architecture, referred to ‘Neuromorphic’ comprising analog synapses and neurons, is a promising alternative to solve the limitation of digital computing, anticipating low power consumption and parallel processing. Recently, emerging non-volatile memories (NVM) such as resistive random access memory (ReRAM), phase change RAM (PCRAM), and ferroelectric transistor (FeFET) have been studied considerably as analog synaptic devices with low-power and high scalability. Research on artificial neuron devices is early-stage compared to synaptic devices, and most literature has demonstrated only proof of concept. Thus, this dissertation focused on developing and investigating nanoscale emerging memory device-based artificial neurons and applying various pattern classification systems. Firstly, by adopting emerging memory devices, integrate and fire neuron (I&F neuron) with energy and area efficiency for perceptron neural network. To investigate the effects of device parameters on I&F neurons with a threshold switching device, I compared neurons using three different TS devices: NbO2 based insulator-to-metal Transition (IMT) device, B-Te based ovonic threshold switching (OTS) device, and Ag/HfO2 based atomic-switching TS device. Furthermore, capacitor-less I&F neuron with Li-based electrochemical random access memory (Li-ECRAM) was proposed, showing linear integration characteristics. Secondly, we proposed a low-power and compact oscillator with NbO2-based IMT device and NVMs for the oscillator neural network. By combination NbO2 and TaOx-based RRAM, I demonstrated a frequency programmable oscillator with 4F2 footprint. In addition, I proposed a linear frequency programmable oscillator with Li-ECRAM, improving the accuracy of ONN-based classification of clustered data.