I am very happy to have completed a 3-month secondment at the University of Amsterdam. This secondment was the first secondment in the framework of CLARIFY. The secondment at UvA was very fulfilling and enjoyable. I was very impressed by the professionalism and depth of research of Dr. Zhao and his research team. Thank you once again.
The secondment at University of Amsterdam consisted of regular technical reports and personal presentations. The topics of the presentations mainly include task scheduling optimization, blockchain application in SLA, anomaly detection in the system, etc. In each group meeting, we discuss with other PhD students about the content of the presentations, the progress of the weekly work, and especially the collaboration on blockchain-related research. Here are some of the presentations that I and others made at the group meeting.
During Secondment, I presented DID-eFed, a system that combines DID with blockchain and federated learning, to demonstrate a potential multi-institutional collaboration scenario.Federated learning is a privacy-preserving collaborative machine learning technique that has been widely used in privacy-conscious branches like healthcare, banking, and insurance. Under the coordination of a central server, each participant used for data trains models locally and periodically uploads them to the central server for secure model aggregation. The aggregated global model has higher accuracy and less bias than the local models and is then distributed back to the participants. Finally, the local model is fine-tuned based on the global model, and the process is iterated until the local model and the global model converge. In this way, the original data will be stored locally, reducing privacy snooping and data misuse due to sharing. In the CLARIFY project, different hospitals and research institutions will use artificial intelligence techniques to mine meaningful information inside medical images.
PhD candidate Hongyun Liu presented me with his research on combining meta learning with reinforcement learning techniques to improve the robustness of task scheduling services in cloud computing systems. His work was later published in the IEEE Cloud Conference.
They designed the meta reinforcement learning based approach to improve the robustness of the scheduling system in dynamic environments and even in new environments. The robustness can be evaluated in two ways: 1) performance deviation after environment changes can be measured as performance loss compared to stable performance. 2) recovery time on adaptation or retraining on new environments or changes. The optimization process of reinforcement learning combined with meta-learning consists of two parts. The outer layer works on learning in different data trajectories to improve the robustness. In the inner layer, the learning goal is to learn the scheduling model, which interacts with the task model and the cloud platform model as a scheduler in the system. The figure below is an introduction to the proposed framework in his paper
. Geng, Jiahui, et al. “DID-eFed: Facilitating Federated Learning as a Service with Decentralized Identities.” Evaluation and Assessment in Software Engineering. 2021. 329-335.
. Liu, Hongyun, et al. “Towards A Robust Meta-Reinforcement Learning-Based Scheduling Framework for Time Critical Tasks in Cloud Environments” IEEE Cloud 2021
Jiahui Geng – ESR3.