Elevator Pitch
We would like to introduce and demonstrate uTensor runtime, a neural network inference library for microcontroller unit (MCU), and its development. We will talk about tensor computing kernels, memory management, CLI, applications and more.
Description
Deploying machine-learning models on resource constrained devices like micro controller units (MCU), require knowledge in both data science and embedded system. However, it’s very hard for developer to master both domains. For example, topics including:
- Knowledge of Machine Learning algorithm such as designing efficient neural network architectures, training dynamic and theoretic/numeric optimization technics
- Memory management, e.g memory allocation/planing
- Coordinate peripheral devices on embedded system
- And more
These have been naturally disjointed domains which makes it difficult for any individual or team to deploy machine learning models on MCUs. With this in mind, uTensor is designed as a developer-friendly neural network inference library targeting MCU.
Neil Tan, the project manager and core developer of uTensor, will talk about the motivation of uTensor and many applications such as sensor fusion, AIoT and robotic that could be powered by uTensor. Kazami Hsieh, another core developer of uTensor runtime, will talk about the design of uTensor runtime and how we make it efficient in terms of memory usage and binary size which are critical for MCU. And Dboy Liao, the core developer of uTensor CLI, will give a demo on how to transforming a pre-trained model using frameworks such as Tensorflow or PyTorch into compilable and runnable uTensor implementation ready for MCU.
Notes
Our agenda is like 10 mins for each of us and 10 mins for the QA section. But it may subject to adjustment, depending on following discussions on the talk within the team.