[Paper Review] Federated Continual Learning with Weighted Inter-client Transfer

Jaehong Yoon, Wonyong Jeong+3 more|arXiv (Cornell University)|Mar 6, 2020

Domain Adaptation and Few-Shot Learning39 references31 citations

TL;DR

FedWeIT decomposes model weights into global, base, and task-adaptive parts, enabling selective inter-client knowledge transfer and sparse communication to improve federated continual learning.

ABSTRACT

There has been a surge of interest in continual learning and federated learning, both of which are important in deep neural networks in real-world scenarios. Yet little research has been done regarding the scenario where each client learns on a sequence of tasks from a private local data stream. This problem of federated continual learning poses new challenges to continual learning, such as utilizing knowledge from other clients, while preventing interference from irrelevant knowledge. To resolve these issues, we propose a novel federated continual learning framework, Federated Weighted Inter-client Transfer (FedWeIT), which decomposes the network weights into global federated parameters and sparse task-specific parameters, and each client receives selective knowledge from other clients by taking a weighted combination of their task-specific parameters. FedWeIT minimizes interference between incompatible tasks, and also allows positive knowledge transfer across clients during learning. We validate our FedWeIT against existing federated learning and continual learning methods under varying degrees of task similarity across clients, and our model significantly outperforms them with a large reduction in the communication cost. Code is available at https://github.com/wyjeong/FedWeIT

Motivation & Objective

• Motivate federated continual learning (FCL) and address interference from irrelevant knowledge across clients.
• Propose a decomposed parameterization to separate global, base, and task-adaptive knowledge.
• Enable selective inter-client knowledge transfer via attention over task-adaptive parameters.
• Improve communication efficiency through sparsity while maintaining or improving task performance.
• Demonstrate superior performance and faster adaptation across varied task similarity scenarios.

Proposed method

• Decompose local models as: θ_c^(t) = B_c^(t) ⊙ m_c^(t) + A_c^(t) + sum_{i≠c} sum_{j<|t|} α_i,j^(t) A_i^(j).
• Use a dense global parameter θ_G derived from aggregating sparsified B_c^(t) ⊙ m_c^(t) across clients.
• Represent knowledge transfer via sparse, attention-weighted aggregation of task-adaptive parameters A from other clients.
• Regularize training with sparsity constraints on masks m_c^(t) and A^(t) and retroactive updates to maintain past task solutions.
• Transmit only sparse, high-impact parameters to minimize communication costs.

Experimental results

Research questions

• RQ1 Can federated continual learning benefit from selectively transferring task-specific knowledge between clients?
• RQ2 How should parameters be decomposed and transmitted to minimize inter-client interference and communication while preserving performance?
• RQ3 Does Attention-based inter-client transfer improve adaptation speed and final accuracy across diverse task similarities?
• RQ4 What is the trade-off between communication cost and accuracy in FedWeIT compared to baselines across multiple datasets?
• RQ5 Is the FedWeIT approach scalable to larger networks and more clients without sacrificing effectiveness?

Key findings

• FedWeIT substantially outperforms single-task, continual learning, and naive FCL baselines across Overlapped-CIFAR-100 and NonIID-50 tasks.
• FedWeIT achieves faster adaptation to new tasks and lower forgetting due to selective cross-client transfer.
• The attention mechanism effectively chooses beneficial task-adaptive parameters from other clients (e.g., matching similar datasets).
• FedWeIT reduces communication costs by transmitting highly sparse task-adaptive and base parameters, while maintaining or improving accuracy.
• Experiments with ResNet-18 show FedWeIT outperforms APD baselines with fewer parameters.
• Across 100 clients, FedWeIT demonstrates strong performance gains and mitigates inter-client interference.