Frameworks, Runtimes & Libraries
7 entriesSoftware that runs trained models on microcontrollers.
Development Platforms & AutoML Tools
8 entriesTools for collecting data, training compact models and generating embedded code.
Microcontrollers, NPUs & Dev Boards
9 entriesHardware used to run TinyML workloads.
Learning Resources & Communities
7 entriesCourses, books and communities for getting started.
Benchmarks, Datasets & Research
5 entriesMeasuring and advancing TinyML performance.
Frequently Asked Questions
What is TinyML?
TinyML is machine learning that runs on microcontrollers and other devices with very little memory and power, often milliwatts or less. Typical uses are keyword spotting, anomaly detection and simple vision on battery-powered sensors.
How is TinyML different from edge AI?
Edge AI covers any AI that runs near the data source, including gateways and GPUs. TinyML is the smallest end of that range, targeting microcontrollers with kilobytes of RAM, so models must be heavily compressed.
How do I get started with TinyML?
Pick a board such as the Arduino Nano 33 BLE Sense, collect sensor data in a platform like Edge Impulse, train a small model and deploy it with TensorFlow Lite Micro. The courses from Harvard and MIT in this directory cover the basics.
How are models made small enough for a microcontroller?
Common techniques are quantization to 8-bit integers, pruning, knowledge distillation and designing architectures for the memory limit, as in the MCUNet research. Inference libraries such as CMSIS-NN then use optimised kernels.
How is TinyML performance measured?
MLPerf Tiny from MLCommons measures latency and energy for tasks such as keyword spotting and anomaly detection on ultra-low-power devices. EEMBC MLMark is another embedded benchmark.