Why On-Device Analytics
Onloading analytics to sensor and edge devices
We have had significant progress in the computing world on offloading computation from client devices to computing infrastructure, such as, from personal computing devices to cloud computing servers. Of late, there is a counter movement, “onloading” computation to the devices that are close to the sources of data. This movement is relevant to sensor and edge devices in the domain of Cyber Physical Systems (CPS) or equivalently Internet of Things (IoT) systems. In this post, we are going to see the reasons behind this movement, what are the success stories so far, and what are the challenges that we need to solve for realizing the full benefit of such computation onloading.
What’s Driving Onloading?
There are three key motivations that are driving this trend. The first is that CPS are leading to massive
increases in the amount of sensory data that we are collecting from our physical environment. This takes the
form of data from smart buildings, smart cities, smart critical infrastructures (like oil and natural gas
pipelines), and digital agriculture farms, among many others. Do we really need all this data to be transferred
to the backend servers, processed there, and stored? Maybe for some applications, the answer is NO for most of
the data. So can we make the decision close to the source of the data which of it can be locally processed and
discarded (or some summary maintained).
The second motivation is the privacy concern. Call it the IMBYO phenomenon, an obvious play on words on the NIMBY (Not In My Back Yard) phenomenon. In IMBYO, we want the data In My Back Yard Only. That is, we do not want the data to leave the premises. Perhaps we would be fine with some digest of the data, or some results of the inference on the data to leave the premises and be handed over to the big vendors who run the cloud farms, but I would certainly not be OK with the detailed data to go out. So I want to onload the computation to the devices on my premises, i.e., devices that I trust.
The third motivation is the application setting whereby the action on the data needs to be taken on the device itself. Thus, the sensory data is processed and leads to some control action on some device that is part of the CPS. For example, the data collected from a digitized farm about the soil nitrate concentration leads to a decision to apply fertilizer through an irrigation system whose nozzle is close to the sensor.