Data-Importance Aware User Scheduling for Communication-Efficient Edge Machine Learning

TL;DR

This paper proposes a data importance-aware user scheduling algorithm for edge learning that considers both communication reliability and data informativeness, leading to faster model convergence.

Contribution

It introduces a novel scheduling method that incorporates data importance metrics based on SNR and data uncertainty for improved edge learning efficiency.

Findings

Faster model convergence compared to traditional schemes.

Effective exploitation of multi-user diversity in communication and data.

Extension of importance metrics from SVM to CNN models.

Abstract

With the prevalence of intelligent mobile applications, edge learning is emerging as a promising technology for powering fast intelligence acquisition for edge devices from distributed data generated at the network edge. One critical task of edge learning is to efficiently utilize the limited radio resource to acquire data samples for model training at an edge server. In this paper, we develop a novel user scheduling algorithm for data acquisition in edge learning, called (data) importance-aware scheduling. A key feature of this scheduling algorithm is that it takes into account the informativeness of data samples, besides communication reliability. Specifically, the scheduling decision is based on a data importance indicator (DII), elegantly incorporating two "important" metrics from communication and learning perspectives, i.e., the signal-to-noise ratio (SNR) and data uncertainty. We first derive an explicit expression for this indicator targeting the classic classifier of support vector machine (SVM), where the uncertainty of a data sample is measured by its distance to the decision boundary. Then, the result is extended to convolutional neural networks (CNN) by replacing the distance based uncertainty measure with the entropy. As demonstrated via experiments using real datasets, the proposed importance-aware scheduling can exploit the two-fold multi-user diversity, namely the diversity in both the multiuser channels and the distributed data samples. This leads to faster model convergence than the conventional scheduling schemes that exploit only a single type of diversity.