Jan Jongboom is an embedded engineer and Developer Evangelist IoT at ARM, always looking for ways to connect more devices to the internet. He has shipped devices, worked on the latest network tech, climbed upon buildings to install gateways and there’s a monument in San Francisco with his name on it. Before he joined the IoT bandwagon he was a core contributor to Firefox OS, and he wrote hundreds of patches to various open source projects.
Talk Title: Machine learning on 1cm2
Machine learning is on top of the hype curve: DeepMind is beating the top Go players, DeepFake generates movies with just Nicolas Cage in it, and self-driving cars have covered over 8 million kilometers. But all these are enabled by throwing copious amounts of computing power at problems, which is not always possible. The cost of raw computing power might be too high, or streaming the raw data is not possible over a low-bandwidth network. Wouldn’t it be great if we could do machine learning directly on small end-devices?
In Arm we’re trying to make this a reality by bringing deep learning directly to microcontrollers. Microcontrollers are small, power efficient and cheap. But they are also slow and have very little RAM. That means we have to be creative! By offloading the training of neural networks to a cluster of large GPUs and only doing classification on the device, we can drastically reduce the computing power needed; and with clever optimizations we can trade a bit of accuracy for a large reduction in memory usage.
In this talk Jan Jongboom will talk about the opportunities for machine learning on microcontrollers, from sensor fusion to auto-encoders, give an introduction to uTensor and CMSIS-NN, and run some real machine learning demo’s in less than 200K of RAM.