Self-supervised learning frameworks for IoT applications

Recent developments in deep learning have motivated the use of deep neural networks in Internet-of-Things (IoT) applications. But the performance of the deep neural network models for IoT applications, like those in other areas, largely depends on the availability of large labeled datasets, which in turn entails significant training costs. Such training costs mainly come from the burden of data labeling whereas the collection of the unlabeled data is a relatively easier process. Motivated by this observation, the self-supervised learning technique which could efficiently utilize the unlabeled data, has been widely studied and got great success in areas of computer vision and natural language processing (NLP). In IoT applications, however, self-supervised learning has not gotten enough attention. Considering the unique characteristics of IoT applications compared with computer vision or NLP tasks, customizing the self-supervised learning frameworks to IoT applications would be an interesting and important topic. Different from computer vision or NLP applications, IoT applications often measure physical phenomena, where the underlying processes (such as acceleration, vibration, or wireless signal propagation) are fundamentally a function of signal frequencies and thus have sparser and more compact representations in the frequency domain. In this dissertation, we build self-supervised learning frameworks for IoT applications by carefully taking the frequency domain characteristics into consideration. We first studied the self-supervised contrastive learning framework, which demonstrated outstanding performance in areas of computer vision or NLP, on IoT applications from a time-domain perspective, and then re-designed the self-supervised contrastive learning framework from the frequency domain. After that, we designed two approaches to: 1) efficiently utilize both of the labeled and unlabeled data, and 2) reduce the dependency on the data augmentation strategies in the self-supervised learning frameworks. We evaluated the performance of our frameworks on public datasets with sensing data like radio signals, measurements of accelerometer and gyroscope, and WiFi signals. Next, we designed a seismic and acoustic signal based target detection system and deployed our self-supervised learning framework on it to study its performance on the real system. In the context of IoT, our customized self-supervised learning frameworks for IoT applications demonstrated obvious performance gain compared with the original ones designed for computer vision and NLP, which shows the effectiveness of our customization.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Liu, Dongxin

Contributors dc:contributor

• Abdelzaher, Tarek
• Caesar, Matthew
• Wang, Gang
• Wang, Peng

• Self-supervised Learning
• Internet-of-Things
• Frequency Domain

Statement dc:rights

• Copyright 2022 Dongxin Liu

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/116199

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F116199