I’m curious if anyone has tried it. This guy seems to think it is possible in theory:
I had a quick scan of the blog - interesting. In my experience with ML, the issue is not the computing mechanism but rather defining and constraining a problem to something that can be solved and then designing a process to automatically undertake that. Thus far my approach has been to collect data from the devices, publish to the cloud, transfer data to Azure Cognitive Services/ Machine Learning Studio and from there refine the parameters that are used to control the process on the device - which are then sent to devices using Particle functions to change the parameters. This works for a small number of devices but due to the manual steps will not for larger numbers of devices. The constraints with the current devices is mainly with RAM to store and manipulate sizeable sets of data. The other point is that MCU devices as their name suggests are controlling and that is their primary aim not statistical processing. My conclusion was that simple ML is possible locally on the MCU - not sure whether this is true ML or statistical Digital Signal Processing.