TinyML: Machine Learning for Microcontrollers
- 1. TinyML: TFLite for Microcontrollers Robert John GDE ML & GCP @robert_thas Machine Learning on Embedded Edge Devices
- 2. Why Machine Learning? Learning from example: activity detection, pet classification ● We can’t write rules for everything ● We produce enough data for our examples ● We have powerful machines for training
- 3. Why Embedded Edge Devices? Some computers do only one thing! ● Placed in hard-to-reach places ● Run on batteries ● Replaced once a year or even longer ● No keyboard or mouse ● Rely on sensors or inputs from other microcontrollers ● Control other microcontrollers or other outputs ● Smart temperature example
- 4. Sensors on Edge Devices More sensors than the average computer ● Motion sensors ● Biometric sensors ● Acoustic sensors ● Environmental sensors ● Force sensors
- 5. Constraints on Edge Devices Can’t eat your cake and have it Raspberry Pi Pico HP Chromebook (Celeron) Raspberry Pi 4 Classification Microcontroller Personal Computer Single-Board Computer Microprocessor RP2040 Celeron BCM2711 Processor Designer Raspberry Intel Broadcom ISA Family Arm Cortex-M0+ IA32 (32-bit x86) Arm Cortex-A72 Clock speed 133MHz 3.5 GHz 1.5 GHz RAM 264 KB 4GB/8GB 2GB/4GB/8GB Storage 2MB Flash HDD/SSD HDD/SSD Inputs GPIO Serial Keyboard/Mouse Keyboard/Mouse Outputs GPIO Serial Monitor/Speakers Monitor/Speakers
- 6. Edge ML Deployment Challenges Inference is all you need! ● Pre-trained models are hard to fit into MCUs ● No OS to act as hardware buffer ● Mobile-optimized models might use instructions not present on MCU ● Need to understand the underlying ISA ● Floating-point arithmetic usually not present ● Model is task specific, no need classify 1000 images! ● Might need to start from scratch!! ● Need to know Python and C/C++
- 7. Keep It Simple! You only need inference! ● The model should only be as large as it needs to be to guaranty the required accuracy ● Leverage quantization-aware training to avoid model degradation ● Implement post-training quantization to shrink models from floating-point (32-bit) to integer (8-bit) ● It’s okay to lose a little bit of context during inference
- 8. Learn More Important Resources ● TensorFlow: https://www.tensorflow.org/ ● TensorFlow Lite: https://www.tensorflow.org/lite ● TFLu: https://www.tensorflow.org/lite/microcontrollers ● TinyML Foundation: https://www.tinyml.org/ ● TinyML Course on EdX: https://www.edx.org/professional-certificate/harvardx-tiny-machine-learning ● Embedded Systems Essentials: https://www.edx.org/professional-certificate/armeducationx-embedded-systems-essentials ●
- 9. Thank You! Robert John GDE ML & GCP @robert_thas