Siavash Mollaebrahim Ghari has followed the PhD Programme at Faculty of Engineering and Science, Specialisation in ICT.
- Trial lecture starts at 10.15
- Public defence starts at 12.15
Title of trial lecture: "Graph Neural Networks: Principles, Architectures, and Applications"
Disputation chair: Morgan Konnestad
Assessment committee:
- First opponent: Professor Antonio Luiz Lopes Ramos, USN
- Second opponent: Associate Professor Naveed ur Rehman, Aarhus University
- Head of assessment committee: Professor Andreas Prinz, UiA
Supervisor in the doctoral work:
- Main supervisor: Professor Baltasar Lozano, UiA
Summary of thesis
Efficient decentralized data processing over networks
Modern technological systems increasingly rely on networks of connected devices that must process large amounts of information. Examples include sensor networks, communication systems, smart infrastructure, and distributed monitoring systems. In many such applications, collecting all available data at one central location can be inefficient, costly, or even impossible. An alternative is decentralized processing, where individual devices cooperate with their neighbours and collectively perform the required computations.
The doctoral dissertation Efficient Decentralized Data Processing over Graphs investigates how such computations can be carried out efficiently using only local communication between connected devices. The network is represented mathematically as a graph, where the nodes represent devices or agents and the connections describe which devices are able to exchange information.
A central problem considered in the dissertation is subspace projection. In simple terms, subspace projection is a mathematical operation that extracts the part of a signal or dataset that belongs to a particular space of interest. It is an important building block in many signal-processing and data-analysis applications. The challenge addressed in this work is how a group of networked devices can collectively perform this operation without requiring a central processor with access to all the data.
The dissertation develops graph-filtering methods that allow the desired computations to be performed through a sequence of local information exchanges. An important objective is to reduce the number of communication rounds required. For symmetric networks, where communication between connected devices is represented symmetrically, the work develops a computationally efficient design that enables exact subspace projection after a finite number of local exchanges and is suitable for larger networks.
The research also considers asymmetric networks, where communication relationships may be directional. Such networks require different mathematical tools because information does not necessarily flow in the same way in both directions. The dissertation investigates the conditions under which exact decentralized subspace projection can be achieved in these networks and develops methods for obtaining accurate approximations when exact implementation is not possible.
A further part of the research addresses a practical challenge: communication networks are not always perfectly reliable. Connections between devices may change or temporarily fail. The dissertation therefore considers decentralized computations over networks with time-varying communication links. Methods are developed in which a sequence of local communication operations collectively implements a desired computation. The work also investigates how information that is unavailable because of communication failures can be estimated, improving the robustness of the distributed processing procedure.
Overall, the dissertation contributes new methods for performing complex computations through local cooperation among networked devices. By reducing dependence on centralized processing and considering symmetric, asymmetric, and time-varying communication networks, the research contributes to the development of decentralized systems that are more efficient, scalable, and robust. These principles are relevant to a broad range of applications in which connected devices must jointly process information while operating under communication and computational constraints.
What to do as an online audience member
The disputation is open to the public. To follow the trial lecture and the public defence online, register on Zoom.
We ask online audience members to join no earlier than 10 minutes in advance. After these times, you can leave and rejoin the meeting at any time.
Opponent ex auditorio:
Deadline for the public to pose questions is during the break between the two opponents. Questions ex auditorio can be submitted to Kristine Evensen Reinfjord.
Contact person
- kristine.reinfjord@uia.no
- Phone
- +47 37 23 30 14