Unsupervised learning of static and dynamic patterns using spike-timing-dependent plasticity of synaptic devices

Abstract

The existing computing system is based on a von-Neumann architecture consisting of a processor and memory, and has continuously improved performance through scaling of the semiconductor device. However, technologies that improve the performance of devices and improve their integration density are increasingly reaching their limits. In addition, since the computing system based on the von-Neumann architecture has a structure that the processor and the memory device are separated, the computing performance is degraded due to latency and high power consumption in the process of data transmission between the processor and memory called the von-Neumann bottleneck. In today’s big data era, it is necessary to recognize and process a lot of data in real time. Therefore, it is necessary to develop a computing system that can handle vast amounts of data with high energy efficiency and at high speed. To this end, efforts are being made to solve the aforementioned problems by imitating the human brain based on how it learns. The human brain is a neural network consisting of 100 billion neurons and 100 trillion synapses. However, with only a small power of 20W, memory, computation, and cognition functions can be processed quickly in parallel. Neuromorphic computing is a technology that mimics the human brain structure and computation process in hardware, enabling low-power, real-time pattern recognition, where related research is being actively conducted worldwide. In this thesis, an unsupervised MNIST pattern recognition simulation using a spike-timing-dependent plasticity algorithm (STDP) using a charge-trap based IGZO TFT and ferroelectric material based synaptic device is conducted to implement neuromorphic computing. Through simulation, we investigate the effects of nonlinearity, asymmetry, cycle-to-cycle, and device-to-device variation of synaptic devices on neural network systems. In addition, the possibility of recognizing not only static patterns but also dynamic patterns is investigated through simulation of object trajectory recognition using dynamic vision sensor imitation data.