Open-source AutoML eases edge AI deployment for developers

An open-source AutoML solution called AutoML for Embedded, co-developed by Analogue Devices and Antmicro, is now available as part of the Kenning framework, aimed at easing the deployment of machine learning models on embedded edge devices.

AutoML for Embedded is designed to streamline and automate many of the typical tasks developers encounter when attempting to implement artificial intelligence on microcontrollers and other resource-constrained hardware. These tasks often include data preprocessing, model selection, hyperparameter tuning, and device-specific optimisation.

Workflow and compatibility

This solution is distributed as a Visual Studio Code plugin and is built upon the Kenning library, emphasising cross-platform compatibility. It integrates with CodeFusion Studio and offers support for ADI's MAX78002 AI Accelerator Microcontroller Units (MCUs) and MAX32690, enabling direct model deployment to these hardware platforms.

The workflow also supports rapid prototyping and testing through Renode-based simulation environments and the Zephyr real-time operating system (RTOS). According to the developers, this flexibility allows users to construct and deploy machine learning models on a wide variety of target platforms, avoiding vendor lock-in.

Step-by-step tutorials, reproducible pipelines, and sample datasets are included to assist users in moving from raw data to edge AI deployment without requiring specialist data science expertise.

Developer-oriented features

The solution is the outcome of collaboration between Analogue Devices and Antmicro, who have combined hardware knowledge with open-source approaches.

"Building on the flexibility of our open-source AI benchmarking and deployment framework, Kenning, we were able to develop an automated flow and VS code plugin that vastly reduces complexity of building optimised edge AI models," said Michael Gielda, Vice President of Business Development at Antmicro. "Enabling workflows based on proven open-source solutions is the backbone of our end-to-end development services that help customers take full control of their product. With flexible simulation using Renode and seamless integration with the highly configurable and standardised Zepher RTOS, the road to transparent and efficient edge AI development using AutoML in Kenning is open."

How the automation works

AutoML for Embedded utilises sequential model-based algorithm configuration (SMAC) to automate the search for optimal model architectures and training parameters. Hyperband with Successive Halving is applied to allocate computational resources towards the most promising candidate models. One of the key features is the automated verification that candidate models will fit within the memory limitations of target devices, allowing for more successful deployment on constrained systems.

After the search and optimisation stages, models can be further refined, evaluated, and benchmarked using standard workflows within the Kenning framework. Detailed reports on model size, inference speed, and accuracy inform user decisions prior to deployment.

Applications and demonstrations

AutoML for Embedded has already been utilised in use cases such as anomaly detection for sensory time series data. In a detailed demonstration, a model created by the tool was deployed on the ADI MAX32690 MCU and tested in both a physical hardware setup and its digital twin using Renode simulation, enabling performance monitoring in real time.

Potential application areas outlined by the project include image classification and object detection on low-power camera systems, predictive maintenance and anomaly detection in industrial IoT sensors, natural language processing for on-device text analysis, and real-time action recognition for sports and robotics settings.

The package is made available to developers via the Visual Studio Code Marketplace and GitHub, reflecting its open-source nature and broad accessibility.