'%3e%3crect%20width='64'%20height='64'%20fill='url(%23fk)'/%3e%3crect%20width='64'%20height='64'%20fill='url(%23bk)'/%3e%3c!--%20Concentric%20radar%20rings%20--%3e%3cpath%20d='M64.0%2014.0A14%2014%200%200%201%2050.0%200.0'%20fill='none'%20stroke='%23fff'%20stroke-opacity='.25'%20stroke-width='1.3'/%3e%3cpath%20d='M64.0%2024.0A24%2024%200%200%201%2040.0%200.0'%20fill='none'%20stroke='%23fff'%20stroke-opacity='.25'%20stroke-width='1.3'/%3e%3cpath%20d='M64.0%2034.0A34%2034%200%200%201%2030.0%200.0'%20fill='none'%20stroke='%23fff'%20stroke-opacity='.25'%20stroke-width='1.3'/%3e%3cpath%20d='M64.0%2044.0A44%2044%200%200%201%2020.0%200.0'%20fill='none'%20stroke='%23fff'%20stroke-opacity='.25'%20stroke-width='1.3'/%3e%3cpath%20d='M64.0%2054.0A54%2054%200%200%201%2010.0%200.0'%20fill='none'%20stroke='%23fff'%20stroke-opacity='.25'%20stroke-width='1.3'/%3e%3c!--%20Thicker,%20high-visibility%20signal%20lime%20arc%20(stroke-width%205.2%20for%20small%20navbar%20display)%20--%3e%3cpath%20d='M54.8%2033.2A34%2034%200%200%201%2030.2%206.5'%20fill='none'%20stroke='%23C8F55B'%20stroke-width='5.2'%20stroke-linecap='round'/%3e%3c/g%3e%3c/svg%3e)
Why do it
- Responds immediately, without a network.
- Uses very little power, so it can run on a battery for a long time.
- Raw data, such as audio, never leaves the device.
Making a model fit
- Choose a small architecture for one narrow task.
- Quantize the model to 8-bit integers.
- Convert it with a runtime built for microcontrollers, such as TensorFlow Lite for Microcontrollers.
- Measure memory, speed and energy on the real chip.
Limits
Models are small and capabilities limited. Updating them in the field is harder than updating a cloud service, so plan for it. Test on the real sensor data and in the real environment, since a model can behave differently away from the lab.
Sources and further reading
- TensorFlow Lite for Microcontrollers. tensorflow.org/lite/microcontrollers.
IICTP is a new independent credential and is not ISO/IEC 17024 accredited.