Universidade Federal do Ceará
Mestrado e Doutorado em Ciências da Computação

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

Data: 30/06/2025
Horário: 14h
Local: Sala de Seminários – Bloco 952

Resumo:

Personalized Federated Learning (PFL) is an extension of the traditional federated learning task that addresses the challenge of data heterogeneity between clients by training a personalized model for each participant. However, personalization and model performance are hampered by the uniqueness and scarcity of data available in each client. In this work, we explore the use of Gaussian processes models coupled with deep kernel learning (DKL) and amortized inference to overcome data scarcity and model performance issues. More specifically, we adapt and evaluate the recently proposed Amortized Variational Deep Kernel Learning (AVDKL) for the PFL setting. We propose to share the DKL component of the model across the clients, while keeping its amortized component for local personalization. Experimental results in distinct learning scenarios indicate that the proposed approach overcomes state-of-the-art GP models in terms of performance, while also presenting faster training times compared to other Bayesian PFL methods.

Banca examinadora:

• Prof. Dr. Javam de Castro Machado (MDCC/UFC) – Orientador
• Prof. Dr. César Lincoln Cavalcante Mattos (UFC) – Coorientador
• Prof. Dr. José Antonio Fernandes de Macêdo (MDCC/UFC)
• Prof. Dr. João Paulo Pordeus Gomes (UFC)