Not necessarily. It may also make bad suggestions because there are gaps in its training (species not included) or because it is drawing incorrect conclusions based on the material it has been provided (e.g. overgeneralizing from habitat or flower preferences). It regularly makes wrong suggestions for bees not because the number of wrongly identified bee observations is exceptionally high (though it is probably higher than in many other taxa), but because several difficult to ID genera and subgenera are only represented in the training set by their untypical members or because it confuses yellow pollen with yellow hair and similar problems.
But this would not be fixed by having it look at every single photo in iNat’s database individually in search of matches. (Not to mention that if the CV were to access the database anew each time someone wanted to use it – which would be the only way to apply it to current data as iNat grows without training it in advance and doing periodic updates – the burden on iNat’s computing power and servers would be astronomical.)
2 Likes
This is the issue with the extremely common moth genus Coleophora. The couple of bright green ones are in the training model, so it’s not trained on the pale streaky ones (which can rarely be ID’d to species). The result is that it throws all the pale streaky Coleophora into a few unrelated bins- specifically Gelechiidae, Plutella, and Crambini. Hunting down where all the Coleophora are being placed at the moment and mass correcting them is a fun pastime. lol
6 Likes
Can I know this detail from those are responsible for this software?
There’s more detail in the FAQ. I don’t claim to know much about software, but from the FAQ pages:
“Computer vision is the process of teaching computers to recognize patterns in images. iNaturalist uses computer vision systems trained on users’ photos and identifications in order to provide automated taxon identification suggestions.”
“The computer vision model is not trained to recognize a certain species—it is trained to recognize typical iNaturalist photos of a certain species.”
https://help.inaturalist.org/en/support/solutions/folders/151000547726
2 Likes
Related to the bee issue, it often gives wrong IDs for planthopper nymphs because one species, Scolypopa australis (native to Australia and a pest in New Zealand) was heavily identified early on. But all the nymphs of that superfamily look very similar, so now the CV tags them all as that species. Every so often I have to go through and zap the ones from around the world that are IDed as that unless they’re represented by adults (which is a bit frustrating because iNat is useful to track the spread of it as an invasive species; it did recently show up for real in Argentina and Hawaii).
4 Likes
Yeah, there are some life stages of some taxa that it doesn’t recognize at all and identifies as something completely different because adults can be identified to species but the larval stage (or whatever) cannot.
This happens occasionally with plants, too, though I think it is less frequent than with insects. Around here, it tends to oversuggest a few species (e.g. Plantago, Reseda) for vegetative rosettes, because they are distinctive and can be identified fairly reliably, while many other dicots are more difficult until they have grown a bit more.
I am also a planthopper identifier and I don’t think I have noticed your profile yet!