A collection of challenges and and stories of what data science looks like in the Internet of Things world, where you need to adapt to the firmware release cycle and have to deal with a low power device that is completely at the mercy of your customers.
The canonical approach of data science is simple: collect some data, train your model, validate and repeat until you get a satisfactory result. Then, stick it into some server and monitor its performance live. But, what happens when your server is actually in your customer’s home and it is a low power device, most of the time offline, that needs to last two years on a pair of AA batteries?
I will present the challenges that we have encountered when developing data science models for the Hive products and our attempts to overcome them. Examples will cover the whole process, from the painful (and obvious) truth that one device can only collect one day of data every day, to translating a machine learning model from scikit-learn to C and fitting that into 32KB of RAM.
Throughout the different phases of development we use the standard python scientific stack: Jupyter Notebook, Pandas, Numpy, Scikit-learn, etc.