Universidade Federal do Ceará
Mestrado e Doutorado em Ciências da Computação

Título: Personalized Federated Learning with Gaussian Processes via Amortized Variational Inference

Data: 20/08/2025
Horário: 14h
Local: Sala de Seminários – Bloco 952 e Google Meet (meet.google.com/frr-bxhn-akg)

Resumo:

Personalized Federated Learning (PFL) extends traditional federated learning by addressing data heterogeneity across clients through the training of customized models for each participant. Despite its advantages, PFL faces challenges due to the limited and non-representative data available on individual clients, which can hinder both personalization and overall model performance. To mitigate these limitations, this work investigates the integration of Gaussian Process (GP) models with Deep Kernel Learning (DKL) and amortized inference within the PFL framework. The recently proposed Amortized Variational Deep Kernel Learning (AVDKL) method is adapted and evaluated in the federated setting. In the proposed approach, the DKL component is shared among clients to learn global representations, while the amortized inference module remains client-specific to enable local personalization. Additionally, the architecture is extended by incorporating a global variational module. The outputs from both the global and personal modules are concatenated to generate hybrid variational parameters, combining global and personalized components. To further enhance this mechanism, a neural network is employed to concatenate the Cholesky decompositions of the two modules, a method referred to as Amortized Cholesky Completion. It is also explored the use of a finetunnig step to the training for better adaption of the personal model to the embedding space from the global model. Experimental results across diverse learning scenarios demonstrate that this approach outperforms existing state-of-the-art GP-based models in terms of predictive accuracy, while also offering faster training times relative to other Bayesian PFL techniques.

Banca examinadora: