Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EcoPull: Sustainable IoT Image Retrieval Empowered by TinyML Models

Mathias Thorsager, Victor Croisfelt, Junya Shiraishi, Petar Popovski

TL;DR

EcoPull introduces a sustainable IoT image retrieval framework by deploying two TinyML models on edge devices: a behavior model that filters out irrelevant images and an image compressor that encodes relevant images into latent representations. A pull-based workflow distributes downlink semantic queries and TinyML models, then collects non-colliding latent transmissions via slotted-ALOHA and reconstructs top results at the receiver. The paper formalizes an energy model and a joint SiFi metric that quantifies both the significance of retrieved data and the perceptual fidelity of reconstructions, and provides analytical and Monte Carlo methods to approximate expected performance. Numerical evaluation using TinyHiFiC-based compression demonstrates energy savings exceeding 70% compared to baselines while preserving retrieval quality, with an optimal compression rate that depends on channel availability and network load. Overall, EcoPull offers a principled, scalable approach to energy-efficient, on-device inference and latent-data transmission for sustainable IoT image retrieval.

Abstract

This paper introduces EcoPull, a sustainable Internet of Things (IoT) framework empowered by tiny machine learning (TinyML) models for fetching images from wireless visual sensor networks. Two types of learnable TinyML models are installed in the IoT devices: i) a behavior model and ii) an image compressor model. The first filters out irrelevant images for the current task, reducing unnecessary transmission and resource competition among the devices. The second allows IoT devices to communicate with the receiver via latent representations of images, reducing communication bandwidth usage. However, integrating learnable modules into IoT devices comes at the cost of increased energy consumption due to inference. The numerical results show that the proposed framework can save > 70% energy compared to the baseline while maintaining the quality of the retrieved images at the ES.

EcoPull: Sustainable IoT Image Retrieval Empowered by TinyML Models

TL;DR

EcoPull introduces a sustainable IoT image retrieval framework by deploying two TinyML models on edge devices: a behavior model that filters out irrelevant images and an image compressor that encodes relevant images into latent representations. A pull-based workflow distributes downlink semantic queries and TinyML models, then collects non-colliding latent transmissions via slotted-ALOHA and reconstructs top results at the receiver. The paper formalizes an energy model and a joint SiFi metric that quantifies both the significance of retrieved data and the perceptual fidelity of reconstructions, and provides analytical and Monte Carlo methods to approximate expected performance. Numerical evaluation using TinyHiFiC-based compression demonstrates energy savings exceeding 70% compared to baselines while preserving retrieval quality, with an optimal compression rate that depends on channel availability and network load. Overall, EcoPull offers a principled, scalable approach to energy-efficient, on-device inference and latent-data transmission for sustainable IoT image retrieval.

Abstract

This paper introduces EcoPull, a sustainable Internet of Things (IoT) framework empowered by tiny machine learning (TinyML) models for fetching images from wireless visual sensor networks. Two types of learnable TinyML models are installed in the IoT devices: i) a behavior model and ii) an image compressor model. The first filters out irrelevant images for the current task, reducing unnecessary transmission and resource competition among the devices. The second allows IoT devices to communicate with the receiver via latent representations of images, reducing communication bandwidth usage. However, integrating learnable modules into IoT devices comes at the cost of increased energy consumption due to inference. The numerical results show that the proposed framework can save > 70% energy compared to the baseline while maintaining the quality of the retrieved images at the ES.
Paper Structure (11 sections, 23 equations, 3 figures)

This paper contains 11 sections, 23 equations, 3 figures.

Table of Contents

  1. Introduction
  2. The EcoPull Framework
  3. Scenario
  4. Description of the Scheme
  5. Energy Consumption
  6. Performance Analysis
  7. Measuring the significance and fidelity of retrieved images
  8. Expected SiFi
  9. Getting approximation of expected
  10. Numerical Evaluation
  11. Conclusions

Figures (3)

  • Figure 1: EcoPull, a sustainable image retrieval framework in which two types of Tiny models are installed into devices: i) a behavior model and ii) an image compressor model.
  • Figure 2: in \ref{['eq:sifi']} against the compression rate $r$ in [bpp]. 'MCMC' refers to the approximation of the expected in \ref{['eq:expected-sifi']}, while 'Simulation' refers to the one obtained by averaging \ref{['eq:sifi']}. The number of available transmission slots $L$ is calculated as $L = {c_{L}} \lceil r^{\mathrm{max}}/{r} \rceil$, where $r^{\mathrm{max}}=4.86$ is the average for the PNG compressed images and $c_{L}$ is a coefficient value thorsager2024generative.
  • Figure 3: Energy saving w.r.t. the baseline $\eta$ as a function of the total number of images per device, $N$.