Improving Global Generalization and Local Personalization for Federated Learning | JoVE Visualize

Area of Science:

Artificial Intelligence
Machine Learning
Data Science

Background:

Federated learning enables collaborative AI model training across decentralized clients while preserving data privacy.
Existing federated learning methods often struggle to balance the creation of a generalized global model with the development of personalized models for individual clients.
Optimizing for one objective (global generalization or local personalization) frequently compromises the other, highlighting a critical limitation in current approaches.

Purpose of the Study:

To introduce a novel federated learning method, personalized federated learning via cross silo prototypical calibration (pFedCSPC), designed to enhance knowledge consistency among clients.
To improve the effectiveness of collaboration between clients by calibrating features from heterogeneous data spaces.
To achieve a better balance between global model generalization and personalized client model performance.

Main Methods:

pFedCSPC utilizes an adaptive aggregation strategy to provide personalized initial models, facilitating rapid adaptation to specific client tasks.
It employs clustering on clients to learn class representation patterns, generating local prototypes that are then aggregated into global prototypes on the server.
A cross-silo prototypical calibration (CSPC) module, using contrastive learning, maps heterogeneous features into a unified space to boost global model generalization.

Main Results:

Experimental results on four datasets demonstrate that pFedCSPC significantly improves both global generalization and local personalization performance.
The method effectively calibrates cross-source features, enhancing the overall collaborative learning process.
pFedCSPC mitigates data imbalance issues and prevents overfitting by leveraging global prototypes to guide local representation learning.

Conclusions:

pFedCSPC successfully addresses the challenge of balancing global generalization and local personalization in federated learning.
The proposed method enhances collaboration effectiveness by unifying heterogeneous features through prototypical calibration.
pFedCSPC offers a robust framework for privacy-preserving collaborative machine learning with improved performance on both generalized and personalized tasks.