I tell you that this was the 1st time ever, I’ve got ID help from users who are working normally outside the Canary area, although earlier I’ve put many comments why I cannot accept a (mostly CV proposed) ID.
It’s after midnight now here, so I should rather go to bed now …
Probably an unpopular opinion, but sometimes when the CV is obviously (but consistently) wrong it’s still helpful. For a long time, the CV would consistently identify galls of Asphondylia ambrosiae as Aceria boycei, despite the two galls looking completely different (the midges’ galls are about the size of a marble and appear alone, and the mites’ galls are the size of cornmeal and appear in large clusters). But it was easy enough to sweep through the Aceria observations and correct the misidentified ones. Most observers would withdraw their incorrect ID if the observation was recent, and I was able to get some help to vote up the older observations. Once we had enough correct IDs, the CV got better, and the problem largely disappeared. Without the consistent misidentification, it would have been hard to find enough observations to retrain the CV.
While it’s true that diligent correction of identifications and more observations of species not in the CV will help improve the overall quality of suggestions in the long term, some of the sources of bad CV suggestions are basically unsolvable without significant changes to the training parameters (e.g.: without incorporating disagreements into the learning process and without finding a way to simultaneously train it on higher-level taxa along with children of those taxa). So I don’t think it’s really helpful to tell people who are getting burned out with the neverending task of correcting CV suggestions that they just need to be more patient.
However, specifically in the cases of endemics with an extremely limited range, I wonder whether it would be feasible to use this range instead of the geomodel for determining where the CV should suggest these species. If I understand correctly, when “endemic” status is assigned to a taxon on iNat, this is associated with a particular place, meaning that there is already information stored in iNat’s database about where it can be expected.
mm I recognise your name because the Canary Islands is included in Africa. For which I ID every day. I have many filters set - including Pre-Mavericks where I will at least pick up IDs which are 2 against 1. For those obs I appreciate when the taxon specialist leaves a brief copypasta - then I can better evaluate whether the 2, or the one Proud Maverick, are correct.
My best example would be Anaxeton arborescens (North of my Fish Hoek gap) and Anaxeton laeve (South of that gap). Effectively two islands divided by the gap, all fitting inside iNat’s 60 km Geomodel hexagon. That needs consistent human effort, started by @botaneek (author of my fynbos field guide). Now easy for me to keep tidy. Collected a few this morning … this obs still shows The Gap till the ID is resolved (or not if the link is removed, but you get the point).
And the exception that proves the rule in Africa ! We HAD cleared the 9 that were wrong.
You might find some of the conversation in this post from a few years ago relevant:
When iNat changed the automated ID suggestion system it made some things less accurate.
It’ll take time and user corrections to fix this.
Since I only have a very rough understanding of how CV works, I would like to know if there is a difference after different reactions to an observer’s incorrect species ID.
- Case A: 1 user responds, disagrees, ID goes back to genus level.
- Case B: 2 or more users reject, same result: Obs. only gets genus level.
Does that make a difference for CV?
Or isn’t it enough if only 1 user destroys the wrong ID?
CV is only trained on observations identified at the level being suggested. So if it’s being trained on Species A, anything you do to kick it out of Species A will have the same impact, ie removes it from the training model. The only time an observation at the genus level is used in the training is if no species within that genus are in the model. Once any species within the genus enters the model, the genus is removed from the training.
It’s also worth noting that the model is “re-trained” and re-released periodically; it’s not learning in real time like most familiar AIs. So anything that’s re-identified now won’t have any effect until the next CV iteration is released.
Not entirely accurate. The most recent model was released last week. At that point, anything with community ID is eligible to be used to train the next version of the model. Anything IDd today will not affect THAT model, but the one after it. @miwi2020, it can be maddeningly slow to see progress, but the model does get better.
In addition to what has been discussed regarding the model and suggestions, another way to reduce erroneous suggestions is to give the model a clean look at the subject. Many CV suggestions are off when the picture is blurry, or with visual clutter, or when the subject is very far away. Cropping can help, but better IDs come from better evidence. Unfortunately identifiers have very little influence on this except to engage regular observers and communicate what views are needed to make a proper ID. I know this has affected how I photograph certain things.
There are plenty of ideas but it looks like none will be implemented.
Search for “the Sorcerer’s Apprentice” on this forum.
The options are:
- Keep on fixing until CV is replaced by something better
- Do something else
Some of us tried fixing a moss genus for over a year. One taxa was added to the model. Based on the current rate of observations, the next taxa would take a couple of years to get added and some over 80 years.
I moved on to the next challenge, something I could achieve during my lifetime.
I am particularly perplexed and scoffing when I see CV suggest something like Juniperus grandis or scopulorum for an observation in Europe or other far flung locations. That suggestion is not driven by observations.
p.s., according to one post, I have the right to complain, though I mainly work on just a couple species of Juniperus.
The CV will suggest “visually similar” and/or “expected nearby” so a user could choose a CV suggestion that is not based on range. You’d have to check the geomodel to see if your place of interest is included or not.
which moss genus was this?
That is not the case in my reality.
If you want to be able to find your comments later (and of course I want that) using: https://www.inaturalist.org/comments?mine=true,
then you cannot put the comment into the ID post
(cannot be found there).
That means you need 2 actions:
- set ID - wait for system response
- set comment - wait for system response
And that can take sometimes up to several seconds in my computer environment, and that multiplies.
Nevertheless in many situations I do set comments.
But regarding rather stupid things like the CV-triggered misIDs of the Micromeria herpyllomorpha I’m kind of fed up sometimes. These users should be happy that they get a reponse at all. And they could always tag me to get an explanation (proud to say that I belong to the users who try to answer every tagging or question). Unfortunately, only a few actually take action themselves. But the ID process is not a one-way street, where I’m supposed to be the service provider who corrects everything and explains it in detail, while the observer has to do almost nothing: one click (CV) at the beginning and one click at the end (agree).
Since the number of botanical observations in the Canary Islands is constantly growing, I simply can’t keep up with editing all the pending posts.
I will increasingly focus on users who take good photos or discover rarer, interesting species.
I can only devote a small portion of my limited time to CV corrections.
Does not seem to be working, though. Observers do not even check if the species being suggested is from the right Continent…
Rosulabryum. (In Australia, Ptychostomum torquescens and Ptychostomum capillare are Rosulabryum as well).
R torquescens is now in the model.