[Paper Review] Scheduling for Cellular Federated Edge Learning with Importance and Channel Awareness

Jinke Ren, Yinghui He+4 more|arXiv (Cornell University)|Apr 1, 2020

Privacy-Preserving Technologies in Data37 references4 citations

TL;DR

This paper proposes a novel probabilistic scheduling policy for cellular federated edge learning that jointly optimizes for both channel quality and the importance of local model updates—measured by gradient divergence—to accelerate convergence and improve learning accuracy. By balancing multiuser diversity in channels and update significance, the method achieves faster model convergence than conventional policies that exploit only one type of diversity.

ABSTRACT

In cellular federated edge learning (FEEL), multiple edge devices holding local data jointly train a neural network by communicating learning updates with an access point without exchanging their data samples. With very limited communication resources, it is beneficial to schedule the most informative local learning updates. In this paper, a novel scheduling policy is proposed to exploit both diversity in multiuser channels and diversity in the "importance" of the edge devices' learning updates. First, a new probabilistic scheduling framework is developed to yield unbiased update aggregation in FEEL. The importance of a local learning update is measured by its gradient divergence. If one edge device is scheduled in each communication round, the scheduling policy is derived in closed form to achieve the optimal trade-off between channel quality and update importance. The probabilistic scheduling framework is then extended to allow scheduling multiple edge devices in each communication round. Numerical results obtained using popular models and learning datasets demonstrate that the proposed scheduling policy can achieve faster model convergence and higher learning accuracy than conventional scheduling policies that only exploit a single type of diversity.

Motivation & Objective

• To address the communication bottleneck in cellular federated edge learning (FEEL) by intelligently scheduling edge devices based on both channel conditions and update importance.
• To overcome the limitation of conventional scheduling that treats all updates equally, regardless of their impact on model convergence.
• To develop a probabilistic scheduling framework that ensures unbiased global model aggregation while maximizing learning efficiency.
• To achieve faster convergence and higher accuracy by jointly exploiting multiuser diversity in wireless channels and diversity in update importance.

Proposed method

• A probabilistic scheduling framework is designed to ensure unbiased model aggregation in FEEL, enabling the selection of updates based on both channel state and gradient divergence.
• The importance of a local update is quantified by its gradient divergence, which reflects its contribution to model improvement.
• For single-device scheduling per round, a closed-form optimal policy is derived to balance channel quality and update importance.
• The framework is extended to support multiple-device scheduling per round, allowing scalable deployment in multiuser cellular environments.
• Theoretical convergence analysis is conducted using strong convexity and Lipschitz continuity assumptions, leading to a convergence rate bound.
• A diminishing learning rate policy is applied to ensure asymptotic convergence, with performance bounds derived via recursive expectation inequalities.

Experimental results

Research questions

• RQ1 How can scheduling in cellular FEEL be improved by jointly considering channel quality and the importance of local updates?
• RQ2 What is the optimal trade-off between channel conditions and update significance in a single-device scheduling scenario?
• RQ3 How does the proposed probabilistic scheduling framework maintain unbiased global model aggregation while prioritizing informative updates?
• RQ4 What performance gains are achievable by exploiting both channel diversity and update importance compared to single-diversity policies?
• RQ5 Can the proposed scheduling policy achieve faster convergence and higher accuracy in practical FEEL settings with real datasets and models?

Key findings

• The proposed scheduling policy achieves faster model convergence and higher learning accuracy than conventional policies that rely solely on channel quality or update importance.
• In single-device scheduling, the closed-form policy achieves the optimal balance between channel gain and gradient divergence, maximizing convergence speed.
• Numerical results on popular deep learning models and datasets demonstrate consistent performance gains across different non-IID data distributions.
• The extension to multi-device scheduling maintains high learning efficiency and scalability in multiuser cellular FEEL environments.
• Theoretical analysis confirms that the method achieves a convergence rate bounded by O(1/T), where T is the number of communication rounds, under diminishing learning rates.
• The method ensures unbiased aggregation and stable convergence even under imperfect channel state information, as validated through rigorous mathematical derivation.