Hardware-Aware AutoML for Edge AI

The AutoML community has made substantial progress in automating model selection, hyperparameter optimization, neural architecture search, and end-to-end machine learning pipelines. However, many AutoML systems still optimize primarily for predictive performance or proxy metrics that do not fully capture the constraints of real-world deployment. Edge AI, and especially microcontroller-based AI, exposes this gap very clearly.

In edge settings, the “best” model is rarely the most accurate model in isolation. A deployable solution must fit into limited flash and RAM, meet real-time latency requirements, respect energy budgets, and remain robust under varying hardware and sensor conditions. These constraints turn model development into an inherently multi-objective optimization problem. They also raise methodological questions that are highly relevant to AutoML:
• Which objectives should be optimized, and which constraints should be treated as hard deployment requirements?
• How reliable are software-based hardware proxies?
• When is hardware-in-the-loop evaluation necessary?
• How can users interpret Pareto fronts and trade-offs between accuracy, latency, memory, and energy?

This tutorial addresses these questions by connecting AutoML concepts with practical Edge AI deployment. Rather than focusing only on one framework or one research contribution, we will introduce the broader design space of hardware-aware AutoML for Edge AI. We will use MIMaaS as a hands-on demonstration platform because it allows participants to upload models, select target devices, and receive real hardware measurements without requiring physical access to microcontrollers. This makes it possible to teach hardware-aware optimization principles in a compact, reproducible, and accessible format.

The tutorial is timely for AutoML 2026 because Edge AI is becoming increasingly relevant across healthcare, industrial sensing (Physical AI), environmental monitoring, audio processing, wearables, and mobile systems. At the same time, the AutoML community is well positioned to contribute methods that make edge deployment more systematic, transparent, and accessible to non-experts.

The tutorial will combine short conceptual blocks with interactive elements. We will use audience questions throughout the session to collect deployment constraints from participants’ own research areas and translate them into AutoML objectives. During the hands-on part, participants will actively inspect hardware reports and make model selection decisions under different hypothetical deployment requirements. Engagement strategies include:
• Short polls on participants’ experience with AutoML, embedded systems, and Edge AI
• Scenario-based questions, such as “Which model would you deploy if energy is the bottleneck?”
• A guided hands-on exercise with prepared models and hardware feedback
• Group interpretation of multi-objective trade-offs
• Discussion of open research challenges relevant to AutoML 2026

The tutorial is designed so that both AutoML researchers and participants with limited embedded-systems experience can contribute meaningfully.

The tutorial targets researchers, students, and practitioners from the AutoML community who are interested in deployment-aware machine learning, multi-objective optimization, neural architecture search, hardware-aware benchmarking, and applied Edge AI. 

The most likely attendees include: 
• AutoML researchers interested in new application domains and deployment-aware optimization 
• PhD students and postdocs working on neural architecture search, hyperparameter optimization, or model compression 
• Researchers in TinyML, embedded machine learning, and hardware-aware AI 
• Applied machine learning researchers who want to deploy models beyond cloud or GPU environments 
• Domain scientists who work with sensor data and need efficient local inference 
• Practitioners interested in practical routes from trained models to microcontroller deployment 

No prior experience with microcontroller programming will be required. Basic familiarity with machine learning and AutoML concepts will be helpful, but the tutorial will introduce the necessary embedded deployment concepts from first principles.