I’ll admit, I’m not really familiar with the intricacies of the AI Training for species identification in iNaturalist.
Macrolepiota colombiana, which a mushroom species originally described from Colombia, and based on an ongoing investigation of the genus, the species, so far, seems to be restricted to Colombia. M. colombiana is also morphologically unique, but because of the lack of study of Macrolepiota in the Neotropical region, most tall and robust species of the genus has been identified as M. colombiana.
My question would then be if the AI is also continuously trained when the observations identified as the species, which probably were used for the AI Training, are being correctly identified.
Thanks in advance,
I don’t know if I am understanding the question correctly, but new computer vision models are released fairly regularly (once a monthish). Species can be included (or not) based on whether there are enough observations or not. So new identifications will change the training set for the model.
There are a fair amount of threads on the CV model on the forum. A couple that might be useful are:
but I’m sure many more. There are also some posts about the CV on iNat itself.
I think the problem you may describing goes beyond simple computer vision issues; There’s a lack of study on Macrolepiota in the Americas (afaik) in general. Actually, a lot of American genuses are just a massive mess (see, Cantharellus, Russula, Neoboletus, Inocybe, Entoloma, etc.) Plus, the CV struggles with mushroom ID in general and generally only seems to do well on extremely distinctive species of fungi.
Couple this with IDers who trust the CV suggestions and may not have much knowledge on mycology, and a lack of experts in certain regions that can even help sort out the mess, and well… you end up with a messy situation.
Best advice is to try to correct what you can and hopefully a comment here and there on posts can help tell people what they should be looking for. CV is going to recommend thing based on the dataset it has, and if the dataset is limited, it can only do so much, and that’s when people have to step in, because ultimately, the CV isn’t even AI, its mostly just pattern recognition
it is the same observation I made on this thread …
For mycology the CV is mostly ‘garbage in - garbage-out’. Correct identification of fungi is often much deeper than macro-morphological visual similarity.
Yes, I’ve been trying to identify most of the observations. And I do understand how difficult the identification of macrofungi based on macromorphological characters is, but obviously, some species are very characteristic and easily recognizable in field based on macromorphology.
My question would be more if the AI (or Computer Vision) actually learns along with the new refined identifications of the species, or rather, if what she were trained on is fixed and can’t be refined.
If I’m understanding how CV works correctly, once the IDs are fixed, it should help fix the CV (especially when the model gets updated, which as noted, happens pretty frequently)
(Though I don’t think it will ever stop trying to ID pieces of trash, piles of leaves, or rotted mushrooms as Grifola frondosa)
About once a month we get an iNat blog post. In there are links to the species ‘added this month’. I asked to have that date added to the taxon page - but tiwane said can’t be done.
It is motivation to achieve the required 100 photos = about 60 obs, for the next CV update.
the photos have to come from research grade observations, right?
RG observations are preferred but others can be used. There is a good thread on this here:
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.