Skip to main content

Fredrika Lundahl: Personalization by combination -The promise and problems of multi-armed bandit orchestrated personalized federated learning

MSc Thesis Presentation

Time: Fri 2021-06-11 09.45

Location: Zoom, meeting ID: 646 0130 8139

Respondent: Fredrika Lundahl

Supervisor: Chun-Biu Li


This thesis is written in collaboration with Ericsson AI Research and is investigating the statistical aspects of a novel, yet unpublished, method to conduct personalized federated learning. The method is the result of a search for a communication efficient personalization algorithm which is also good at handling poisonous workers. Poisonous workers here refers to workers containing faulty data, created on purpose or by mistake. The proposed model originates from the Federated Averaging algorithm but replaces the averaging over the models of all (or most) workers by the averaging over the models from a selected combination of workers only. The combination is selected by a multi-armed bandit inspired decision rule, which learns the appropriate set of workers to combine in order to optimize the performance on a given target worker. Consequently, the decision rule also learns which workers to avoid. The model is the most advantageous when the number of workers is small and each worker has a small dataset compared to the complexity of the model. One of the biggest differences compared to other personalization methods is that the proposed method personalizes only for one worker at a time instead of for all workers simultaneously. This enables personalization in situations where other methods struggle, such as when the workers contain little data.

The biggest challenges with the proposed methods arise from the fact that the training of the decision rule happens during the training of the global federated learning model. This causes problems for the training of the decision rule. One problem is the lack of samples from the reward distributions, the training period can be very limited as it decided by the federated learning model and not the decision rule. Another problem is the variance coming from the fact that the global model is changing every round. We propose some ways to combat these challenges. A limitation of the model is that it cannot handle many workers, as the number of possible combinations of workers becomes too many. By incorporating cluster analysis in future works, we might be able to reduce the number of combinations to consider and aggregate models with greater care. This should give the proposed model wider applications as it then should be able to personalize even when the number of workers is big and each worker contains more data.

Belongs to: Department of Mathematics
Last changed: Jun 04, 2021