Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Energy-Efficient Edge Learning via Joint Data Deepening-and-Prefetching

Sujin Kook, Won-Yong Shin, Seong-Lyun Kim, Seung-Woo Ko

TL;DR

This paper tackles energy-limited edge learning by enabling IoT devices to offload high-dimensional data efficiently to nearby edge servers. It introduces Joint Data Deepening-and-Prefetching (JD2P), which orders features via PCA, progressively offloads only what is needed (data depth/deepening) and strategically prefetches future features to extend the transmission window. Threshold designs for both Binary SVM and Multi-class DNN classifiers (MoCs: distance to hyperplane, negative entropy, and posterior gap) govern depth decisions, while a closed-form optimal prefetching policy minimizes expected energy. Experiments on MNIST demonstrate substantial energy reductions with negligible impact on accuracy, and the framework is extensible to online and federated edge learning, with potential integration into future 5G/6G MEC systems.

Abstract

The vision of pervasive artificial intelligence (AI) services can be realized by training an AI model on time using real-time data collected by internet of things (IoT) devices. To this end, IoT devices require offloading their data to an edge server in proximity. However, transmitting high-dimensional and voluminous data from energy-constrained IoT devices poses a significant challenge. To address this limitation, we propose a novel offloading architecture, called joint data deepening-and-prefetching (JD2P), which is feature-by-feature offloading comprising two key techniques. The first one is data deepening, where each data sample's features are sequentially offloaded in the order of importance determined by the data embedding technique such as principle component analysis (PCA). Offloading is terminated once the already transmitted features are sufficient for accurate data classification, resulting in a reduction in the amount of transmitted data. The criteria to offload data are derived for binary and multi-class classifiers, which are designed based on support vector machine (SVM) and deep neural network (DNN), respectively. The second one is data prefetching, where some features potentially required in the future are offloaded in advance, thus achieving high efficiency via precise prediction and parameter optimization. We evaluate the effectiveness of JD2P through experiments using the MNIST dataset, and the results demonstrate its significant reduction in expected energy consumption compared to several benchmarks without degrading learning accuracy.

Energy-Efficient Edge Learning via Joint Data Deepening-and-Prefetching

TL;DR

This paper tackles energy-limited edge learning by enabling IoT devices to offload high-dimensional data efficiently to nearby edge servers. It introduces Joint Data Deepening-and-Prefetching (JD2P), which orders features via PCA, progressively offloads only what is needed (data depth/deepening) and strategically prefetches future features to extend the transmission window. Threshold designs for both Binary SVM and Multi-class DNN classifiers (MoCs: distance to hyperplane, negative entropy, and posterior gap) govern depth decisions, while a closed-form optimal prefetching policy minimizes expected energy. Experiments on MNIST demonstrate substantial energy reductions with negligible impact on accuracy, and the framework is extensible to online and federated edge learning, with potential integration into future 5G/6G MEC systems.

Abstract

The vision of pervasive artificial intelligence (AI) services can be realized by training an AI model on time using real-time data collected by internet of things (IoT) devices. To this end, IoT devices require offloading their data to an edge server in proximity. However, transmitting high-dimensional and voluminous data from energy-constrained IoT devices poses a significant challenge. To address this limitation, we propose a novel offloading architecture, called joint data deepening-and-prefetching (JD2P), which is feature-by-feature offloading comprising two key techniques. The first one is data deepening, where each data sample's features are sequentially offloaded in the order of importance determined by the data embedding technique such as principle component analysis (PCA). Offloading is terminated once the already transmitted features are sufficient for accurate data classification, resulting in a reduction in the amount of transmitted data. The criteria to offload data are derived for binary and multi-class classifiers, which are designed based on support vector machine (SVM) and deep neural network (DNN), respectively. The second one is data prefetching, where some features potentially required in the future are offloaded in advance, thus achieving high efficiency via precise prediction and parameter optimization. We evaluate the effectiveness of JD2P through experiments using the MNIST dataset, and the results demonstrate its significant reduction in expected energy consumption compared to several benchmarks without degrading learning accuracy.
Paper Structure (38 sections, 1 theorem, 33 equations, 14 figures, 2 algorithms)

This paper contains 38 sections, 1 theorem, 33 equations, 14 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Prior Work
  3. Contributions and Organizations
  4. System Model
  5. Edge Learning Scenario
  6. Data Embedding
  7. Offloading Model
  8. Joint Data Deepening-and-Prefetching: Overview
  9. Data Depth
  10. Data Deepening
  11. Data Prefetching
  12. Metrics of Clarity
  13. Binary SVM classifier
  14. Multi-class DNN classifier
  15. Hierarchical Edge Inference
  16. ...and 23 more sections

Key Result

Proposition 1

Given the ratio of prefetching $\rho_k$ in round $k$, the optimal prefetching data size $p_k^*$, which is the solution to Problem:upper bound, is where $\varphi = \left(h_k\nu \right)^{\frac{1}{\ell-1}}\frac{\tau_k}{t_{k+1}}$ and $s_k$ is the number of data samples in $\mathbb{S}^{(k)}$.

Figures (14)

  • Figure 1: The edge learning network comprising IoT devices and an edge server collocated with a wireless access point.
  • Figure 2: An example of data deepening from the $1$-dimensional space to the $3$-dimensional space. The data samples in the gray area are included in the ACS set. Only data samples in the ACS take into account offloading following features marked by dotted lines in the subsequent rounds.
  • Figure 3: The data prefetching architecture. Each round comprises offloading duration, training duration, and feedback duration. The $(k+1)$-th ACS set, $\mathbb{S}^{(k+1)}$, can be available when the round $(k+1)$ starts. A few data samples' $(k+1)$-th features, $\mathbf{x}_{m,k+1}$, are prefetched from the IoT device to the edge server during the training duration of the $k$-th depth classifier. The remaining data samples' $(k+1)$-th features are offloaded after training if they belong to the $(k+1)$-th ACS set.
  • Figure 4: The ACS region in the 1-dimensional space is obtained by the probability distribution and distance from the hyperplane.
  • Figure 5: The number of prefetching data obtained by Proposition \ref{['proposition 1']} is equal to the number of optimal prefetching data. Parameters are set $s_k = 1000$, $\ell = 3$ and $\tau_k = t_{k+1} = 0.5$.
  • ...and 9 more figures

Theorems & Definitions (8)

  • Definition 1: Data Depth
  • Remark 1: Data Importance and Feature Importance
  • Remark 2: Symmetric ACS Region
  • Remark 3: Communication Efficiency vs. Learning Accuracy
  • Proposition 1: Optimal Prefetching Policy
  • Remark 4: Effect of Parameters
  • Remark 5: Effect of Parameters
  • Remark 6: Channel Capacity vs. Energy Consumption