Abstract
Recent advances in neural network computing and cloud computing have enabled the integration of artificial neural networks (ANNs) and the internet of things (IoTs), thereby introducing artificial intelligence of things (AIoTs). As neural network computing requires high computing power and large resources, the majority of AIoT neural network computing relies on cloud computing, which performs energy- and data-intensive computation on a remote server with enormous computing resources. However, AIoT device developments have enabled the integration of multiple, high-quality sensors, resulting in the generation of huge amounts of data, which causes delays in the data transfer between local devices and remote servers. Furthermore, the transfer of the original raw data collected by AIoT devices to a remote cloud server has raised privacy concerns during the data transfer required for using cloud computing. Therefore, the demand for computing on local devices, as opposed to cloud computing alone, has increased.
Recently, edge computing has attracted significant interest as a solution for the aforementioned problems. Edge computing is a computing framework that brings data-intensive computing closer to data sources such as sensors in the AIoT devices for a faster response time. To enable edge computing, the hardware on local devices must provide energy-efficient and fast computing. Here, this dissertation focuses on demonstrating edge computing by integrating sensors with in-memory computing units to perform neural network algorithm computation on local devices, thereby enabling instant computation on sensors with low energy consumption. The following chapters introduce the components required for edge computing, such as optoelectronic and pressure sensors, non-volatile memories, oxide semiconductor transistors, and neuromorphic computing. This chapter concludes with a summary of the entire dissertation.
