ECML PKDD 2021 ML meets IoT Tutorial Part I: ML for IoT Devices
0 likes99 views
This document discusses machine learning for Internet of Things (IoT) devices. It begins by noting the rapid growth of networked devices and importance of edge computing. It then discusses the hardware capabilities and limitations of various edge devices from microcontrollers to more powerful boards. The document introduces TensorFlow Lite for microcontrollers and model optimization techniques. It demonstrates running a TensorFlow Lite model on an Arduino board. Finally, it discusses several ML applications for resource-constrained edge devices like hearing aids and medical mask detection.
ECML PKDD 2021 ML meets IoT Tutorial Part I: ML for IoT Devices
- 1. Machine Learning Meets Internet of Things: From Theory to Practice Part I: ML for IoT Devices Bharath Sudharsan ECML PKDD 2021 Tutorial
- 2. IoT Edge Computing Valuable extractable knowledge at the edge ▪ Cisco annual Internet report ✓ By 2023, there will be more than three times more networked devices on Earth than humans ✓ About half of the global connections will be machine-to-machine (M2M) connections ▪ Push applications, data and computing power to the edge of the Internet ▪ Deploying, managing apps outside the data centers, on the customer-owned devices
- 3. IoT Edge Computing - Hardware View ARM Cortex-M0 MCU based BLE beacon Edge gateway with GPUs and SSDs Powerful CPU + basic GPU based SBCs (single board computers) Edge computing hardware: highly resource constrained -> high resource (left to right) ▪ MCUs and small CPUs: BLE beacons, smart bulbs, smart plugs, TV remotes, fitness bands ▪ SBCs: Raspberry Pis, BeagleBones, NVIDIA Jetsons, Latte Pandas, Intel NUCs, Google Coral ▪ GPU accelerated: AWS snowball, Digi gateways, Dell Edge Gateways for IoT, HPE Edgeline ▪ Why MCU? Roughly 50 billion MCU chips were shipped in 2020 (market estimates), which far exceeds other chips like GPUs & CPUs (only 100 million units sold)
- 4. IoT Edge Computing - Hardware Spec ▪ No file system, Lack Python support (only C, C++), only few hundred MHz clock speed, low SRAM and Flash memory ▪ No parallel execution units), lack inbuilt hardware accelerators such as APU (Accelerated Processing Unit), KPU (convolution operation accelerator), FPU (Floating-point accelerator), and FFT (Fourier transform accelerator) Popular open-source MCU boards with specification
- 5. TensorFlow Lite for Microcontrollers
- 6. TensorFlow Hub ▪ Models ready to use from off the shelf ▪ Optimizers applicable: Quantization, Pruning, etc. ▪ Can add additional tunable layers at the network end
- 7. Netron NN Model Visualizer ▪ Demo: Download popular models and visualize and brief analysis using https://netron.app/ ✓ yolo-v5-tflite from https://tfhub.dev/neso613/lite-model/yolo-v5-tflite/tflite_model/1 ✓ 3 types of fully connected Neural Networks (NNs) from https://github.com/bharathsudharsan/TinyML-Benchmark-NNs-on-MCUs FC 1 x 10: 1 layer with 10 neurons FC 10 + 50: 2 layers, where 1st layer contains 10 neurons, and 2nd layer contains 50 neurons FC 10 x 10: 10 layers, where each layer contains 10 neurons
- 8. Arduino IDE + TF Lite Library ▪ Install the Arduino TensorFlow Lite library ▪ Run the Arduino TF Lite hello world example ✓ Files -> Examples -> Arduino_TensorFlowLite -> hello_world. The example sketch should appear ✓ Upload on MCU, open the Serial Plotter, observe LED ✓ Model trained to replicate sine function. Generates data pattern to either blink LEDs or control an animation
- 9. Dev Boards - Hardware View ▪ Pinout summary ✓ Power ✓ Logic ✓ I2C & SPI ✓ GPIO ✓ Analog
- 10. Dev Boards - Hardware View ▪ ST-LINK V2 programmer provides a more reliable way to flash the Blue Pill boards ▪ Full range of STM32 SWD interface debugging ▪ The tiniest Arduino ATtiny85 is flashed (programmed burned) using UART IC
- 11. Dev Boards - ML Applications https://cartesiam.ai/product/
- 12. Dev Boards - ML Applications
- 13. Dev Boards - ML Applications https://www.edgeimpulse.com/blog
- 14. Conclusion ▪ A small device with an enormous impact on our lives. For examples ✓ TinyLSTMs: Efficient Neural Speech Enhancement for Hearing Aids ✓ A Tiny CNN Architecture for Medical Face Mask Detection for Resource-Constrained Endpoints
- 15. Contact: Bharath Sudharsan Email: bharath.sudharsan@insight-centre.org www.confirm.ie