Personalized Federated Learning with Hierarchical Two-Branch Aggregation for Few-Shot Scenarios.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Personalized federated learning (pFL) trains client-specific models to handle data heterogeneity.
• Few-shot learning presents challenges in pFL due to overlooked neural representation hierarchies and rigid inductive biases.
• Existing methods are vulnerable to distribution shifts in federated environments.

Purpose of the Study:

• To propose pFedH2A, a hierarchical framework with brain-inspired mechanisms for few-shot personalized federated learning.
• To address limitations in balancing generalization and personalization under few-shot conditions.
• To overcome vulnerability to distribution shifts in federated learning.

Main Methods:

• Designed a dual-branch hypernetwork (DHN) to generate aggregation weights, mimicking the brain's perceptual and representational processing for fine-grained personalization.
• Introduced a relation-aware module to learn adaptive similarity functions for clients, enabling few-shot classification without rigid prototype assumptions.
• Utilized hierarchical structure and brain-inspired mechanisms for improved federated learning.

Main Results:

• pFedH2A demonstrated superior performance compared to existing pFL baselines in few-shot scenarios.
• The dual-branch hypernetwork effectively captured both shared and personalized representations.
• The relation-aware module provided adaptive similarity measures crucial for few-shot classification.

Conclusions:

• pFedH2A offers an effective solution for personalized federated learning in few-shot settings.
• The framework's brain-inspired mechanisms enhance the ability to balance generalization and personalization.
• The proposed methods show promise in addressing data heterogeneity and distribution shifts in federated environments.