Dynamic Task Clustering and Aggregation for Decentralized Federated Multitask Learning

Federated Multi-Task Learning (FMTL) enables privacy-preserving collaboration among clients optimizing di”erent but related tasks, yet most existing approaches rely on centralized servers for coordination, creating single points of failure and additional privacy and security exposure. This thesis presents a fully decentralized peer-to-peer FMTL framework in which clients infer task relatedness from observable training signals, select neighbors dynamically, and aggregate updates without central coordination. Across datasets with di”erent task-correlation structures, the evaluation reveals a consistent conclusion: the optimal aggregation scope depends on task correlation strength. For strongly correlated tasks, selective sharing (e.g., backbone-only) can be beneficial, whereas weakly correlated settings often require broader aggregation to maintain stable convergence. The thesis further addresses practical numerical instability in conflict-aware aggregation through a stabilization mechanism and reports failure cases to delineate safe operating regimes. Overall, the work demonstrates server-free FMTL for complex dense prediction workloads and provides actionable guidance for similarity-based collaboration and aggregation design.