New Computer Vision Model (v2.17) with over 1,000 new species!
We've released v2.17 today with 1,145 new species (95,903 taxa up from 94,758). This new model was trained on data exported on September 29th, 2024.
Here's a graph of the model's release schedule since early 2022 (segments extend from data export date to model release date) and how the number of species included in each model has increased over time.
The graph below shows model accuracy estimates using 1,000 random Research Grade observations in each group not seen during training time. The paired bars below compare average accuracy of model 2.16 with the new model 2.17. Each bar shows the accuracy from Computer Vision alone (dark green) and Computer Vision + Geo (green). Overall the average accuracy of 2.17 is 87.5% (statistically the same as 2.16 at 87.6% - as described here we probably expect ~2% variance all other things being equal among experiments).
Here is a sample of new species added to v2.17:
Comments
Excellent, well done...and still no additional delays in responsetime ? I am always wondering if the there is more demand in computing power if the model changes/increases in size...or longer response times.
For a genus like Russula, where red Russulas are basically indistinguishable via photo alone (see https://www.mushroomexpert.com/russula.html for a humorous summary), I wish they would be removed from the computer vision model- the model does not have enough vision to distinguish red Russula species like R. silvestris shown in this update.
@hannadv you can add disagreeing genus level IDs to any of these observations if you don't think they can be confidently ID'd as R. silvestris https://www.inaturalist.org/observations/identify?reviewed=any&quality_grade=needs_id%2Cresearch&verifiable=true&taxon_id=49447
the model is only as strong as the dataset its trained on, so thanks for all the work weeding this garden!
Thanks @loarie I try to do so (at least in my region) but when the computer vision tempts people with a species ID that cannot be confirmed, the temptation usually wins! Gotta keep up on #identiFriday to keep those species IDs at bay.
The addition of Piper methysticum is worthy of celebration out in my corner of the planet! A good time to tip a coconut cup of pounded rootstock of this culturally significant Pacific island plant! Awesome!
@loarie I think what @hannadv is pointing out is that while the observations that are teaching the computer vision may be correct (many of the R. silvestris observations that are posted are confirmed by DNA tests), the problem is that future observations of red Russula mushrooms must also be DNA verified for any certainty. Since visual characteristics are not enough information for suggesting a species, then maybe the computer vision should not suggest any species at all—only genus. This would be true of some other fungi genera as well. Thoughts?
Palila!
I'm no longer a CV sceptic. The accuracy of the iNat CV model is impressive.
@dsmorris @hannadv this is a relatively common situation with plants and invertebrates as well, where some species in a genus (or larger group) can be identified while other species are never possible to identify. The CV recognizes the impossible species as being similar to the possible species, and recommends the wrong species because it doesn't know about the existence of the impossible species. There are lists on the forum (1, 2) of species for which identifiers need to constantly keep up with these erroneous CV identifications.
It can help a lot if there are multiple possible but very similar species that the CV knows (let's say a genus of 5 species, where 2 are possible to ID but challenging, while the other 3 are impossible to ID), because then it gets less confident between them and goes back to genus to be safe.
Add a Comment