Spatial Dependency in Edge-native Artificial Intelligence
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
Lecture hall L3, Linnanmaa campus, remote connection: https://oulu.zoom.us/j/61903666579
Topic of the dissertation
Spatial Dependency in Edge-native Artificial Intelligence
Doctoral candidate
Master of Science Lauri Lovén
Faculty and unit
University of Oulu Graduate School, Faculty of Information Technology and Electrical Engineering, Center for Ubiquitous Computing
Subject of study
Computer Science and Engineering
Opponent
Professor Schahram Dustdar, Vienna University of Technology
Custos
Docent Susanna Pirttikangas, University of Oulu
Locality in AI methods and applications, distributed in small devices and the network infrastructure
Edge computing augments cloud computing. While cloud computing is based on far away computing centres, edge computing acknowledges the computing resources in the continuum between local devices and the cloud. The computing resources in edge computing are often heterogeneous, with varying capacity, intermittent connectivity, and opportunistic availability. In contrast, modern artificial intelligence and especially machine learning methods are often deployed in the cloud, and assume the computing resources are homogeneous, abundant, centralized, easily scalable, and always available.
This thesis studies edge AI, a nascent field of research combining edge computing and artificial intelligence. A particular focus in the thesis is on spatial dependencies, which quantify the similarity of observations in the spatial dimension. Spatial dependencies are prominent in edge AI due to the local nature of edge service users, the computational resources, as well as many of the observed data-generating processes. The thesis asks three research questions. The first one looks for a method to explicitly consider spatial dependencies in edge AI, while the second and third ones apply the method for edge server placement and environmental sensing.
As result, the thesis first proposes a novel spatial clustering method, named PACK, which partitions a set of spatial data points according to configurable attributes and constraints. PACK then provides a basis for edge server placement and workload allocation, where a large-scale edge deployment can be optimized such that user quality of experience and deployment quality of service are maximised. Furthermore, PACK serves a crucial function in environmental sensing with a massive fleet of mobile sensors, providing grounds for distributing computations and data for a novel, edge-native method for interpolation. In both edge server placement and environmental sensing, the proposed methods outperform state-of-the-art. Finally, the thesis looks at the limitations of the proposed methods, their significance, and maps potential avenues for future research.
This thesis studies edge AI, a nascent field of research combining edge computing and artificial intelligence. A particular focus in the thesis is on spatial dependencies, which quantify the similarity of observations in the spatial dimension. Spatial dependencies are prominent in edge AI due to the local nature of edge service users, the computational resources, as well as many of the observed data-generating processes. The thesis asks three research questions. The first one looks for a method to explicitly consider spatial dependencies in edge AI, while the second and third ones apply the method for edge server placement and environmental sensing.
As result, the thesis first proposes a novel spatial clustering method, named PACK, which partitions a set of spatial data points according to configurable attributes and constraints. PACK then provides a basis for edge server placement and workload allocation, where a large-scale edge deployment can be optimized such that user quality of experience and deployment quality of service are maximised. Furthermore, PACK serves a crucial function in environmental sensing with a massive fleet of mobile sensors, providing grounds for distributing computations and data for a novel, edge-native method for interpolation. In both edge server placement and environmental sensing, the proposed methods outperform state-of-the-art. Finally, the thesis looks at the limitations of the proposed methods, their significance, and maps potential avenues for future research.
Last updated: 1.3.2023