Lol. Yeah, not quite Skynet though.
There are actually quite a few useful applications that may not be obvious. Many of which we already have solutions for but a machine learning approach can actually provide improved results in some respects.
Take a “fuzzy” CO2 controller. This is a closed loop system that “learns” how to apply CO2 depending on how the data is changing based on the interaction between a variable or two. ML is applied in a similar manner but the number of inputs can be much more varied and not directly obvious to how they would affect CO2, for instance. Surprising insights are discovered in this way. For instance, the controllers we have today look at the rate of CO2 change and perhaps account for exhaust events. Taking this one step further, for instance, if we integrate the PAR information into the machine we glean more information about how the CO2 will change apriori to changes in light intensity. That is before the change to the environment actually occurs (on things we normally take for granted). The machine figures out how such changes will affect CO2 based on a bunch of loosely connected pieces of data. Sometimes it’ll guess the correct course of action even if the machine has never encountered a unique situation.
Or how about the leaf image processing. Detection of nutrient deficiencies can be performed using normal image processing and recognition (a form of ML) but then further expanded to account for data collected on other environmental considerations. With a large enough dataset, you end-up being able to “automatically” optimize a system even under a constantly changing environment or for unexpected/unknown situations. Cool stuff could result…
Overkill, perhaps. But, this stuff is coming to you “soon” and it’ll likely be packaged in the same footprint as our current controllers…