An Experimental Study of Split-Learning TinyML on Ultra-Low-Power Edge/IoT Nodes

2025-07-22

AI Summary

Deploying deep learning models on ultra-low-power edge devices—such as microcontrollers—is hindered by severe memory and computational constraints. Method: This paper introduces the first wireless performance evaluation framework for Split Learning tailored to TinyML, implemented on an end-to-end testbed based on the ESP32-S3. It enables over-the-air benchmarking across multiple protocols—ESP-NOW, BLE, UDP/IP, and TCP/IP—for intermediate activation transmission. Using 8-bit quantized MobileNetV2, it measures latency and energy consumption for transmitting 5.66 kB activation tensors. Results: UDP achieves 3.2 ms one-way latency and a steady-state RTT of 5.8 s; ESP-NOW reduces RTT to 3.7 s; BLE significantly extends device battery lifetime. The study systematically characterizes the critical trade-off between latency and energy across wireless protocols, providing empirical evidence and design guidelines for deploying Split Learning in real-world IoT environments.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Application Domains: Internet of Things, Sensor Networks & Smart Cities
• Search and Optimization: Learning to Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Energy management for devices in mobile Web and WoT environments
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Running deep learning inference directly on ultra-low-power edge/IoT nodes has been limited by the tight memory and compute budgets of microcontrollers. Split learning (SL) addresses this limitation in which it executes part of the inference process on the sensor and off-loads the remainder to a companion device. In the context of constrained devices and the related impact of low-power, over-the-air transport protocols, the performance of split learning remains largely unexplored. TO the best of our knowledge, this paper presents the first end-to-end TinyML + SL testbed built on Espressif ESP32-S3 boards, designed to benchmark the over-the-air performance of split learning TinyML in edge/IoT environments. We benchmark the performance of a MobileNetV2 image recognition model, which is quantized to 8-bit integers, partitioned, and delivered to the nodes via over-the-air updates. The intermediate activations are exchanged through different wireless communication methods: ESP-NOW, BLE, and traditional UDP/IP and TCP/IP, enabling a head-to-head comparison on identical hardware. Measurements show that splitting the model after block_16_project_BN layer generates a 5.66 kB tensor that traverses the link in 3.2 ms, when UDP is used, achieving a steady-state round-trip latency of 5.8 s. ESP-NOW presents the most favorable RTT performance 3.7 s; BLE extends battery life further but increases latency beyond 10s.

Problem

• Evaluates split-learning TinyML on ultra-low-power edge/IoT nodes
• Benchmarks wireless protocols for split-learning performance
• Assesses memory-compute trade-offs in microcontroller-based deep learning

Innovation

• Split-learning TinyML on ESP32-S3 boards
• Quantized 8-bit MobileNetV2 model partitioning
• Wireless activation exchange via ESP-NOW/BLE/UDP/TCP