“How Do We Enable Edge ML Everywhere? Data, Reliability and Silicon Flexibility,” a Presentation from Edge Impulse
0 likes103 views
The document discusses enabling edge machine learning (ML) for real-time applications, emphasizing improved manufacturing productivity and efficiency through advanced visual inspection systems. It highlights challenges in scaling edge ML, such as the necessity for data-centric ML approaches and optimal performance calibration on various silicon hardware. Additionally, it showcases advancements in hardware and ML algorithms that enhance object detection capabilities in constrained environments.
“How Do We Enable Edge ML Everywhere? Data, Reliability and Silicon Flexibility,” a Presentation from Edge Impulse
- 1. How Do We Enable Edge ML Everywhere? Data, Reliability, and Silicon Flexibility Zach Shelby Co-founder and CEO Edge Impulse
- 2. Visual inspection system to monitor worker safety and flag delays on the production line in real-time. Advantech increases manufacturing productivity by 15% ● A reported 15% overall increase in production line efficiency ● Faster detection of idle time raises assembly-line productivity ● Managers free up time to focus on production planning and operations 2 © 2022 Edge Impulse
- 3. State of the edge ML 2022 3 © 2022 Edge Impulse
- 4. What are the barriers to edge ML scale? 4 © 2022 Edge Impulse
- 6. Big data, big model is a problem 6 Dataset Test Deploy Model • State of the art: Monolithic batch, from big data to big model • Developed for ML research, unsuitable for most edge ML applications © 2022 Edge Impulse
- 7. Data-Centric ML – from zero to hero 7 Test Deploy Anomaly • Auto-clustering, using feature analysis to help experts quickly label data • Auto-labeling, using pre-trained expert models to suggest labels • Active learning, using inference to drive the data collection process Model Dataset Auto-clustering Auto-labeling Active Learning © 2022 Edge Impulse
- 8. Auto-clustering in action 8 © 2022 Edge Impulse
- 9. Making ML industrial grade
- 10. Sensor fusion for industrial-grade ML Image Processing Sensor Audio Image ML Classifier © 2022 Edge Impulse 10
- 11. Sensor fusion for industrial-grade ML Image Processing Sensor Audio Image ML Classifier DSP Features Post- processing ML Classifier Anomaly Detector © 2022 Edge Impulse 11
- 12. Sensor fusion for industrial-grade ML Image Processing Sensor Audio Image ML Classifier DSP Features Multi-modal ML Classifier Anomaly Detector Post- processing Audio Features © 2022 Edge Impulse 12
- 13. • Model training validation != performance • Requires testing on the entire algorithm with real-world data for realistic performance • Understand the impact of post-processing while accounting for device constraints and latency • Choose the ideal balance between false activations and false rejections • Leverage genetic algorithms to design optimal post-processing configuration Calibrating performance at scale © 2022 Edge Impulse
- 14. ML on today and tomorrow’s silicon
- 15. Hardware is sexy again! 15 Sensor Audio Image Video MCU + FPU CPU GPU High-end MCU Low-end MCU RA MCU PSoC 6 © 2022 Edge Impulse
- 16. Hardware is sexy again! 16 MCU + FPU CPU GPU High-end MCU Low-end MCU Sensor Audio Image Video Wakeup NPU NDP100 RA MCU PSoC 6 © 2022 Edge Impulse
- 17. Hardware is sexy again! 17 MCU + FPU CPU GPU MCU + NPU Low-end MCU Sensor Audio Image Video Wakeup NPU NDP100 NDP200 Akida M55/U55 Katana RA MCU PSoC 6 © 2022 Edge Impulse xG24
- 18. Hardware is sexy again! 18 MCU + FPU CPU + NPU NPU + GPU MCU + NPU Low-end MCU Sensor Audio Image Video TD4VA Wakeup NPU NDP200 RZ MPU NDP100 RA MCU PSoC 6 © 2022 Edge Impulse Akida M55/U55 Katana xG24
- 19. • The power of hardware profiling • Digital twin of ML on hardware • We are combining hardware profiling with hyperparameter search – EON Tuner • Hardware-aware AutoML across data, pre- processing and ML blocks Hardware profiling & tuning 19 © 2022 Edge Impulse
- 20. FOMO: Faster objects, more objects ● Object detection on constrained compute ● 20x performance improvement ● Scales down to 100k RAM to enable MCUs ● Better at detecting smaller and more numerous objects ● Capable of segmentation and counting objects Cortex-M4 Cortex-M7 Cortex-A Nvidia FOMO 2 fps 15-30 fps 60+ fps 150+ fps SSD NA NA 3 fps 20 fps © 2022 Edge Impulse 20
- 21. Resources Increasing Manufacturing Efficiency https://casestudies.edgeimpulse.com/industrial-iot-with-advantech Constrained object detection: FOMO Blog docs.edgeimpulse.com/docs/tutorials/fomo-object-detection-for- constrained-devices Request a demo edgeimpulse.com/schedule-a-demo 21 © 2022 Edge Impulse
- 22. Don’t miss these talks! FOMO: Real-time Object Detection on Microcontrollers Jan Jongboom, Co-founder and CTO, Edge Impulse • Date: Wednesday, May 18 • Start Time: 10:50 am Deep Dive: Develop and Deploy Advanced Edge Computer Vision—Fast! Jenny Plunkett and Shawn Hymel, Snr. DevRel engineers • Date: Thursday, May 19 • Time: 9 am – 12:00 pm 22 © 2022 Edge Impulse