TinyAirNet: TinyML Model Transmission for Energy-efficient Image Retrieval from IoT Devices

TL;DR

The paper addresses energy-efficient pull-based data collection for IoT by preemptively transmitting a TinyML feature extractor to edge devices to enable on-device semantic querying. It introduces TinyAirNet, a three-phase framework that decomposes total energy into $E_{\mathrm{comp}}$ and $E_{\mathrm{comm}}$, and provides analytic expressions for total energy $E_{\mathrm{total}}$ and retrieval accuracy $\gamma$ through phases that include threshold-based data selection via $V_{\mathrm{th}}$. Simulations using EtinyNet1.0 parameters demonstrate significant energy savings, up to $67\%$, under accuracy constraints as the number of stored images rises, outperforming a baseline offloading scheme for large datasets. The work argues for generalizing preemptive TinyML model transmission to a broader class of data-collection tasks, informing energy-aware semantic communication design in 6G IoT.

Abstract

This letter introduces an energy-efficient pull-based data collection framework for Internet of Things (IoT) devices that use Tiny Machine Learning (TinyML) to interpret data queries. A TinyML model is transmitted from the edge server to the IoT devices. The devices employ the model to facilitate the subsequent semantic queries. This reduces the transmission of irrelevant data, but receiving the ML model and its processing at the IoT devices consume additional energy. We consider the specific instance of image retrieval in a single device scenario and investigate the gain brought by the proposed scheme in terms of energy efficiency and retrieval accuracy, while considering the cost of computation and communication, as well as memory constraints. Numerical evaluation shows that, compared to a baseline scheme, the proposed scheme reaches up to 67% energy reduction under the accuracy constraint when many images are stored. Although focused on image retrieval, our analysis is indicative of a broader set of communication scenarios in which the preemptive transmission of an ML model can increase communication efficiency.

Figures (3)

• Figure 1: An example of for -empowered data collection.
• Figure 2: Total energy consumption and retrieval accuracy of against the threshold of $V_{\mathrm{th}}$.
• Figure 3: $V_{\mathrm{th}}$ and $\eta$ against $N$.