Addressing Realisms Faced by Deep Learning Models in Cyber Physical Systems
In recent years, applications in the cyber physical systems (CPS) area have greatly benefited from deep learning's success. However, there exists an intrinsic problem with directly applying a trained deep learning model on a CPS application: CPS applications have constraints arising from realisms, whereas the training of deep learning models often does not take realisms into consideration. Among the multitude of realisms, there are four types of realisms that are most important. The first realism is task-specific, resulting from the interaction between the deep learning model and the environment in which the deep learning model is deployed. The second realism is non-targeted samples. The third realism is that many data-driven deep learning models are not robust against even minor changes. The fourth realism is the realism that comes during the process when the deep learning model is deployed. In this thesis, we show how to deal with these four realisms using specific applications as demonstrations. We design specific algorithms for the applications, each of which is extensively evaluated on samples that have realisms within them. In addition, we demonstrate how to deal with the realisms using deployments in people's homes, each of which lasts for three to four months.
John Stankovic, Advisor (CS/SEAS/UVA)
Yangfeng Ji, Chair, (CS/SEAS/UVA)
Hongning Wang (CS/SEAS/UVA)
Lu Feng (CS, ESE/SEAS/UVA)
Homa Alemzadeh (CS/ECE/UVA)