Edge AI and TinyML

45 contact hours of lecture / lab courses, 60 hours of preparation/follow-up for the lab course, 45 hours of follow-up for the lecture and exam preparation

Basic programming skills (Python, C/C++); basic knowledge in the field of Machine Learning / Deep Learning; basic knowledge of microcontroller technology and computer architectures; experience in using software development tools (IDE, CLI, debugging).

Students acquire knowledge about methods, tools, and hardware platforms for developing, optimizing, and deploying AI applications on edge systems. They understand the differences between Edge AI on more powerful embedded platforms and TinyML on extremely resource-constrained microcontrollers.

Upon successful completion of the module, students will be able to:

• explain the fundamentals of Edge AI and TinyML, categorize their areas of application, and evaluate the use of edge, cloud, and hybrid inference with respect to latency, energy efficiency, data privacy, and resource requirements,
• analyze the characteristics of different edge hardware platforms and select appropriate hardware and software solutions for specific application scenarios,
• specifically optimize machine learning models for resource-constrained edge systems using quantization, pruning, clustering, and knowledge distillation,
• develop embedded AI applications, implement the preprocessing of sensor data, and deploy optimized models on edge devices and microcontrollers, • measure, analyze, and evaluate the performance of edge AI applications with regard to accuracy, latency, memory requirements, and energy efficiency, and derive appropriate optimization measures.

Fundamentals of Edge AI and TinyML: Terms, concepts, differentiation from cloud AI, and fields of application; comparison of architectures; evaluation based on latency, energy efficiency, data privacy, and resource requirements

Edge Hardware Platforms: Microcontrollers, NPUs/TPUs, single-board computers; comparison of processing power, memory, and energy consumption

Software Ecosystems and Toolchains: Frameworks (e.g., TensorFlow Lite Micro, Edge Impulse, CMSIS-NN, ONNX Runtime), development workflow

Model Optimization: Quantization, pruning, clustering, knowledge distillation; analysis regarding size, accuracy, and speed

Sensor Data Processing: Acquisition and preprocessing (audio, image, acceleration), feature extraction under resource constraints

Development and Deployment: Model training, conversion, integration; deployment on microcontrollers/edge devices

Performance Analysis: Accuracy, latency, memory, energy consumption; benchmarking methods; derivation of optimization measures

Blackboard, slides, projector, instructional/educational videos, guest lectures

• Pete Warden, Daniel Situnayake: TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers, O’Reilly, 2019.
• Vijay Janapa Reddi: Introduction to Machine Learning Systems, MIT Press, 2026.
• Other online resources and publications