I actually think this is a great application for AI. “Hey AI, look at these 2 million trees and tell me which ones look similar to this one.”
It identifies a bunch of trees and even though it identifies a bunch that aren’t similar then you have a much smaller list to sort through. It doesn’t have to be right all the time, it’s just helping narrow it down.
This is way different from “Hey AI, what time does the show start tonight”, where if you get it wrong 10% of the time then it’s a useless tool.
AI is great where statistical accuracy is more valuable.
This would be a good test to run competitive models over, one model is optimized to find the target tree with a 80+% confidence, the second model is optimized to find all trees which are not the target with the same confidence. Where the two models agree, run the first model again but with a confidence requirement of 99+% (which will take much longer to run) over the smaller data set.
Unfortunately it doesn’t quite work that way. The dataset they are training it on contains images of a single tree, so it’s ability to generalise to a normal tree of that species will be incredibly limited.
Consider a facial recognition algorithm trained only on images of Nicolas Cage, then being tasked with identifying members of his family. It would do very well at identifying Nicolas Cage in a crowd, but probably not a good job of identifying anyone else.
Would it help if you photoshopped a bunch of trees with different superficial characteristics but kept the defining traits of the subspecies and trained it on those images?
Maybe, if you could reliably render known traits based on descriptions for which we likely don’t have photographic evidence.
You risk tainting the model though. If some artefact of the photoshop gets detected well by the model, then it will quickly learn to identify photoshopped trees, not trees that actually look like the target species.
Ah that makes sense. Kind of like the old AI problem where it thought fish had fingers because most of the training material had people holding up the fish.
I actually think this is a great application for AI. “Hey AI, look at these 2 million trees and tell me which ones look similar to this one.”
It identifies a bunch of trees and even though it identifies a bunch that aren’t similar then you have a much smaller list to sort through. It doesn’t have to be right all the time, it’s just helping narrow it down.
This is way different from “Hey AI, what time does the show start tonight”, where if you get it wrong 10% of the time then it’s a useless tool.
So basically AI is good at categorizing data where perfect accuracy isn’t needed, but manual categorization isn’t feasible.
Kinda,
AI is great where statistical accuracy is more valuable.
This would be a good test to run competitive models over, one model is optimized to find the target tree with a 80+% confidence, the second model is optimized to find all trees which are not the target with the same confidence. Where the two models agree, run the first model again but with a confidence requirement of 99+% (which will take much longer to run) over the smaller data set.
Unfortunately it doesn’t quite work that way. The dataset they are training it on contains images of a single tree, so it’s ability to generalise to a normal tree of that species will be incredibly limited.
Consider a facial recognition algorithm trained only on images of Nicolas Cage, then being tasked with identifying members of his family. It would do very well at identifying Nicolas Cage in a crowd, but probably not a good job of identifying anyone else.
Would it help if you photoshopped a bunch of trees with different superficial characteristics but kept the defining traits of the subspecies and trained it on those images?
Maybe, if you could reliably render known traits based on descriptions for which we likely don’t have photographic evidence.
You risk tainting the model though. If some artefact of the photoshop gets detected well by the model, then it will quickly learn to identify photoshopped trees, not trees that actually look like the target species.
Ah that makes sense. Kind of like the old AI problem where it thought fish had fingers because most of the training material had people holding up the fish.
Yes you make a good point. Perhaps they trained it on other trees of a similar family?