Title: Machine Intelligence of Ubiquitous Computing in the Internet of Things
Ph.D. Candidate: Fei Dou
Major Advisor: Dr. Jinbo Bi
Associate Advisors: Dr. Song Han, Dr. Sheida Nabavi, Dr. Chun-Hsi Huang (external)
Committee Members: Dr. Derek Aguiar, Dr. Suining He
Date/Time: Tuesday, May 23rd, 2023, 9:30 am
Location: WebEx and In-Person
In-Person Location: HBL 1102
Meeting link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m230c8bff562c17715403ad630a6caaad
Meeting number: 2624 595 0123
Join by video system: Dial email@example.com
You can also dial 184.108.40.206 and enter your meeting number.
Join by phone: +1-415-655-0002 US Toll
Access code: 2624 595 0123
The penetration of technologies such as Machine Learning (ML), Artificial Intelligence (AI), wireless broadband, and the Internet of Things (IoT) is propelling the rapid adoption of ubiquitous devices across a variety of sectors. However, the enhancement of machine intelligence in ubiquitous computing in the IoT is hindered by various barriers, including: 1) inefficiency and low scalability of trained models, 2) security and privacy concerns surrounding user data, 3) heterogeneity of data across devices, with imbalanced data distribution on individual devices, 4) communication bottlenecks and high computational expenses.
Three interconnected research thrusts are investigated in this dissertation research by developing new methods from the perspectives of reinforcement learning, federated learning, and contrastive learning. Firstly, we design a bisection reinforcement learning approach using a novel Markov Decision Process (MDP) formulation for indoor localization to ensure the efficiency and effectiveness of model training in large solution spaces, and to improve the scalability of trained models. Secondly, we propose an on-device ILBS framework by developing a personalized federated reinforcement learning method to rigorously protect the privacy of personal data on IoT devices over wireless edge networks, while addressing model oscillation and reducing communication costs in the presence of system heterogeneity. Lastly, we develop a latent orthonormal contrastive learning approach using high-resolution paired remote sensing imagery to address model drift under small batch size and asymmetrical configuration with imbalanced data, promoting more discriminative classification and reducing memory burden.