New Generation Federated Learning.

Area of Science:

• Machine Learning
• Distributed Systems
• Internet of Things

Background:

• Federated learning (FL) is a distributed machine learning (ML) approach enabling training without data exchange, crucial for the Internet of Things (IoT).
• Existing FL frameworks assume fixed tasks and unlimited storage, which is unrealistic due to evolving client classes and storage limitations.
• Real-world applications require FL to handle dynamic task sequences and strict storage constraints.

Purpose of the Study:

• To introduce a novel federated learning framework, New Generation Federated Learning (NGFL), that addresses the limitations of current FL setups.
• To integrate incremental learning principles into FL to manage continuous learning and dynamic task arrivals.
• To provide a rigorous mathematical representation and analyze the challenges of this new framework.

Main Methods:

• Developed a new federated learning framework (NGFL) incorporating incremental learning principles.
• Defined a mathematical representation for the NGFL framework, considering continuous private task learning on clients.
• Identified and addressed key challenges: aggregation of heterogeneous output layers and task transformation mutual knowledge.

Main Results:

• Established lower and upper baselines for the proposed NGFL framework.
• Demonstrated the framework's capability to handle dynamic task sequences and storage limitations.
• Addressed critical issues in combining incremental learning with federated learning.

Conclusions:

• NGFL is a more desirable and practical framework for real-world federated learning scenarios, especially in evolving IoT environments.
• The integration of incremental learning is essential for adapting FL to dynamic and resource-constrained settings.
• Further research on addressing the identified challenges will enhance the robustness and applicability of NGFL.