iNat has the computer vision to help identify stuff based on photos. i wonder if there would be a way to have some sort of AI look at the past identification notes and comments for various taxa as a way of offering up text suggestions for how to identify?
for example, i’ll often go through and correct identifications for Rudbeckia amplexicaulis and offer up Rudbeckia hirta or subsection Rudbeckia as a correction, noting that R. amplexicaulis would not have hairy green parts. could an AI look at those past identifications / comments and tell someone wanting to select R. amplexicaulis from the computer vision suggestions that R. amplexicaulis does not have hairy green parts?
It’s hard to see how that could be implemented, but I think what you are doing is helpful. If you consistently go through and correct out IDs for Rudbeckia with hairy green parts, the CV will catch on to that sooner or later. I have seen the CV suggestions for “my” taxa improve considerably over time.
And how to misidentify. The ML is garbage for the critters I am interested in, and while I understand it may be better for other taxa, having some deranged bot providing handy hints still would be worse than Clippy with a PhD. I’d like an off button, therefore.
I agree with russellclarke, for fungi the CV model is often less than useless, requiring a lot of human input to correct. And because of the nature of fungi, I doubt the visual model will EVER become genuinely useful given all of the additional qualitative information that needs to be input to come to an identification. The CV can’t smell the mushroom, for example.
There is no substitute for human knowledge and understanding in these areas, not now and possibly not ever, and putting all of our eggs in the AI basket rather than the public education basket seems foolish and shortsighted.
Smell is subjective and is simply an example. If you look at projects like this one with incredibly limited scope, it is an incredible amount of work to create something that will input characteristics and output a species ID. https://www.nature.com/articles/s41598-021-92237-5
It was designed for a single genus, something like 11 species iirc, and required more than 10 data points, including spore dimensions(!) in order to produce accurate results.
Pictures will not ever come close to cutting it for fungi and there is probably never going to be a standard set of data that one can input for an observation that will output a useful result.
Correcting erroneous IDs very obviously generated by the CV model is a huge waste of fungal identifiers’ time (and other Kingdoms as well I’m sure). If the user instead had to do actual work (online research, field guides, etc) they may come up with something resembling the ‘right answer’.
If there were enough iders who’d help observers with fungi, the latter would lern when it’s possible to rely on cv and when it’s not, cv can id quite a few species, but can’t a lot more, but it’s 0 attendance from iders that leads to spreading of wrong ids and wrong ideas in peoples’ heads. It’s impossible to expect everyone to make a research on every shroom they meet when they can see hundreds of them each day and no means to learn species if they only get cv ids.
I don’t think having an AI, using human expert ids, and helping observers to learn is something mutually exclusive. On the contrary: it would be great to have a conversational field guide AI which asks the right questions and guides the observer to help with the identification (and at the same time teach the observer how to ID these particular organisms, and maybe even to produce good training data to make the AI better). Example: the observer takes a picture of a mushroom cap from above, AI suggests to take another from below, after that to make a spore print / cut the stipe / look for mycorrhizal association / smell etc. This is obviously science fiction right now, but with the progress of AI technology AND the help of identifiers this could be reality some day.
I hear you. But it is part of iNat reaching from all the way up to - the scientist who has rewritten the taxonomy of … down to an iNatter like me who offers Fungi … and waits to see what mushroom people can do with the obs.
You can force the ID up - then mark DQA as good as it can be. Extra work, but some scientists do that for their own dedicated corner. Also the taxon pictures - I’m not a mushroom person - but there is one with a cut stem which turns blue - that is clearly diagnostic. Making taxon pictures useful is also pre-emptive.
Better onboarding would corral observers into IDing at the level where they are confident. Instead of accepting with goodwill when iNat says - pretty confident this plant is a black bear - which I got offered yesterday (we don’t have any bears except teddies here)