Running AI/Neural networks on microcontrollers made simple with the STM32Cube.AI
summary
STM32Cube.AI converts pretrained Keras/Caffe/TensorFlow models into optimized C code that calls STM32 libraries and reports which MCU parts the model fits. It supports quantization/fixed point and TensorFlow Lite / TensorFlow for Micro, with ONNX/MXNet/PyTorch coming next year. Examples: a wearable activity model runs with ~2 KB RAM, 12 KB flash and <1 mA (1 inference/sec); a food-vision MobileNet variant needs ~200 KB RAM and delivers 5–6 FPS on an M7 at 400 MHz. Hardware: SensorTile, IOT node, OpenMV and evaluation boards.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:51Artificial Intelligence (AI)
- 3:55Al Edge Computing (Embedded)
- 5:19Neural Networks on STM32 Simple, fast, optimized
- 5:36The Key Steps Behind Neural Networks Neural Network INN Model Creation
- 7:12Al / Neural Networks on Microcontrollers
- 8:39ST Toolbox for Neural Networks
- 14:38Form Factor Hardware to Capture and Process Data
- 15:16Fast Go to Market Module to Capture Data with More Accuracy
- 15:53Form Factor Hardware Allot Node for More Connectivity
- 16:20Distributed Al: Sensor + STM32 Optimize Performance and Power Consumption
- 17:28Human Activity Recognition (HAR) Motion Example in FP-AI-SENSING1 Package
- 18:42Human Activity Recognition IGN (5 classes)
- 19:31Image Classification Vision Example in FP-AI-VISION1 Package
related talks
04mMultiple embedded and Cloud AI solutions at ST
32mAI for Mobile and IoT Devices: TensorFlow Lite
19mWhy Machine Learning on Microcontrollers?
27mMachine Learning on the edge using TensorFlow Lite
55mEdgeAI - AI in the Palm of Your Hand
55mBuilding with TensorFlow Lite for microcontrollers
21mThings I learned running neural networks on microcontrollers
34mTensorFlow Lite