Embedded AI Podcast E18: Shawn Hymel on Edge AI - NPUs, Deployment Challenges, and the Future of ML

We talk with Shawn Hymel about edge AI and running machine learning on microcontrollers. Shawn creates embedded AI courses and helps engineers navigate this fast-changing field. We cover keyword spotting in smart speakers to today's NPUs handling full object detection on tiny devices. We discuss privacy benefits of on-device processing—your doorbell doesn't send everything to the cloud—and the reality of programming these systems. Shawn is honest about challenges: vendor lock-in, dependency hell, and days spent getting inference running. But capabilities are remarkable: YOLO object detection at milliwatt power on microcontrollers. We discuss what embedded engineers need to learn (statistics, probability, knowing when NOT to use AI), and where the field is heading. Reinforcement learning makes robots walk on unpredictable terrain. Neuromorphic computing with spiking neural networks might beat GPUs for efficient AI. But sometimes, a PID loop is still the right answer. 0:00 Welcome to Embedded AI Podcast 2:36 Why Run ML on Microcontrollers 5:01 Keyword Spotting Killer App 6:15 NPUs Enable Object Detection 11:04 Privacy Through On-Device Processing 13:25 Programming NPUs Under the Hood 19:14 Development Pain Points and Challenges 24:20 AI Transforms Course Creation 28:02 Training Embedded AI Engineers 39:53 When Not to Use AI 43:05 Reinforcement Learning for Robotics 46:56 Neuromorphic Computing Future 50:52 Resources and Closing Remarks