In the fourth iteration of the course, students continue to produce increasingly impressive projects. And somehow there were even more projects. As Edge AI continues to develop and gain traction in the wild (i.e., outside academia), the software, hardware, and tools available for embedded ML are becoming more mature. These developments are helping students in the course be more creative in their projects.
Here are some of my impressions about how developments in Edge AI affected student projects:
- Hardware development: More powerful microcontrollers are available in small form factors and with low power consumption, offering larger SRAM and better support for peripherals such as high-resolution cameras.
- Software development: Students used a variety of new approaches. These included:
- Small object detection models, such as FOMO or YOLO Pro, which can fit on more powerful microcontrollers.
- Sequential LSTM (Long Short-Term Memory) models for embedded applications.
- Newly developed on-device or small LLMs (Large Language Models) for generative AI.
- Knowledge distillation techniques to create soft class probabilities for small models.
- Tools development: Students had access for the first time to Edge Impulse Enterprise Edition, which offers AI labeling of datasets and a higher compute time limit.
Some of the earlier course challenges remained. These include:
- Real-time programming for embedded systems and RTOS (Real-Time Operating Systems).
- Multicore programming and synchronization.
- Programming for ML hardware accelerators, such as neural decision processors.
Needless to say, when the students gave their final demo, I was super pleased and proud of their projects and the course’s overall direction. Please check out the students’ project for Fall 26 here.
Please check out the student projects for Fall 2025!
