Auto-IDs of gastropods are completely useless

Auto-IDs with snails simply didn’t work at all, some snails of the different subclasses have almost identical shell, but it keeps auto-suggesting IDs with remark “we’re pretty sure this is…”. Most recent one:
https://www.inaturalist.org/observations/59674501
Correct ID even not in the list and it’s “pretty sure” that this is an American genus in Europe! It’s not a singular case, some ridiculous examples (few of many):
https://www.inaturalist.org/observations/49211750
https://www.inaturalist.org/observations/13579976 (suggesting different subclass)
https://www.inaturalist.org/observations/52498770
It keeps suggesting American Helicina orbiculata for nearly all European Stylommatophora (different subclass). What should be done is that auto-IDs for gastropods should NEVER contain “we’re pretty sure this is…” and I would rather even add additional notification about usual “top suggestions” that they usually didn’t work with snails

Few more examples:
https://www.inaturalist.org/observations/50402663
https://www.inaturalist.org/observations/52597774
https://www.inaturalist.org/observations/54035786

Hi @igor117 and welcome to the Forum!
I think you might want to join in on the conversation at https://forum.inaturalist.org/t/computer-vision-clean-up-wiki/7281

5 Likes

Thanks, well, I’ve added a comment there, but I don’t think that it can be fixed by fixing some separate species

1 Like

I want toguard AI a little, AI confuses snails a lot, but still very often shows right ids in suggestions for common species, and it learns a lot each year, plus e.g. suggesting Helicoidea for many-many species it can’t name for sure.
Some examples shown for you are tricky, so AI works as good as it can now, it’s in our hands to teach it those species.
But in the end people shouldn’t rely on AI when iding groups they’re not familiar with.

3 Likes

Well, when it suggests Helicoidea for Pomatias elegans that not belongs to it, but to different subclass of gastropods, it’s not that much useful. I’m not saying that auto-suggestions should be switched off for gastropods, but when it chooses one taxon and says “we’re pretty sure this is it” or “мы почти уверены, что это относится к вот этому” more people will put erroneous IDs instead of simply “Gastropoda” or “Mollusca”. As a result there are tons of ridiculous IDs with American species in Asia etc. It should learn also to analyze distribution to work with gastropods. Many species of gastropods can be reliable identified only by the genitals even to the family level if to discard distribution data and when ID should be done by very few characters on the photo an only way to do it is to choose between the species recorded from this region in literature.

Can you explain more about what you mean by these suggestions? If you have an actionable feature request, please submit a topic to the #feature-requests category (if one doesn’t already exist).

3 Likes

The simplest thing that I suggest is to made it not auto-choosing one “top suggestion” with remark “we’re pretty sure this is it” for all suggestions for gastropods and only left usual list of suggestions. If more complicated, perhaps something could be added to this list of suggestions like “please notice that auto-suggestions of IDs for snails has very low accuracy”

What we need to do is adding lots of right ids, AI couldn’t id European plants the way it does now just 1,5 years ago, if there will be thousands of good RG observations we will have much higher % of good suggestions, you can check how good it learns insect genera even in groups of many similar ones, but it needs lots of info to learn the next time of system “reloading”.
You can check multiple threads about adding notifications and how AI is using location, etc. gastropods are not different than other hard groups, maybe you’d like to spend your vote on those feature requests.

6 Likes

Most of the plants can be ID by good photo, but 90% of gastropods, if to choose from the world fauna, only can be ID by the genitals after dissection or by the minute structures of the shells that cannot be visible on some random photo. So no, it will never work good with gastropods unless it will start analyzing distribution data, no matter how good it learning, as it’s simply impossible by the random photo only. Same with many Diptera and Myriapoda where even best experts simply cannot ID by the photo because they need anatomy or some very small structures.

Genera and families can be ided, and you can see it in Diptera and Myriapoda suggestions, changing the way AI chooses the head suggestion will change nothing, people who rely solely on AI can pick whatever they find fitting, and it’s not always the first suggestion.
iNat is work of people 1st, not machine and not what new users choose, @bernhardhausdorf and @kevin474 spend their time on gastropods and probably can add something from their workflow experience.
AI chooses american species as there’re millions of obs from USA, not because it’s working so bad. Adding distribution won’t change anything if you believe AI can’t see the small difference between gastropod species, it still will be wrong. I find it controversial and I see how it changes over time for hard groups. When I find american species’ id in Europe I check them all and clear, if I keep it clear then AI won’t add “seen nearby” to future suggestions thus there will be less wrong picks.

5 Likes

That’s the thing that I’m talking about that some land snails from the different families of 3 different subclasses can have almost identical shells that even experts would not be able to recognize by random photo, e.g. Helicinidae (Neritimorpha), Cyclophoridae (Caenogastropoda) and many families of Stylommatophora - Helicidae, Camaenidae, Bradybaenidae etc (Heterobranchia). So how AI will learn to ID these families if it’s not possible by photo only?

Why do you keep referring to “random photos”? What is it? If an expert can’t id it, then no AI changing will help. Once leaving a comment to a person that it’s crucial to id to the level of confidence and that it’s impossible to id the group from the photo - they will learn, keep it saved and just copy-paste, it does help, every interaction makes difference.

6 Likes

As I said before, an expert can choose between species of some region and ID by the limited number of characters from the limited number of species. Without knowing location (i.e. choosing from the world fauna) only few most recognizable snails and slugs can be ID and AI is choosing from the world fauna by far. E.g. in Eastern Europe there are little over 200 species of land gastropods and most of the species can be ID (but still not all) if I know these 200 species and that photo was taken in Eastern Europe. In the world there are over 30000 species of land gastropods many of which are very similar, but live in the different parts of the world, so even if to know all these 30000 species, without knowing location, only a few of them can be ID by the photo with general view of a snail or slug.

What you say was said many times before, it’s understandable, and there should be a request for changing AI geography, thing is it’s a massive amount of work to do for millions of species.

2 Likes

Two critical (maybe three) things here, at least in my opinion:

  1. AI suggestions are just that, suggestions. They are not meant to be definitive IDs, they are meant to get users in the ballpark and to be followed up by people who have more experience and can confirm, refine, or correct said ID.

  2. If you want better AI IDs, then spend more time providing identifications to other observations and getting them to research grade. That’s what the AI learns from. There are lots of genera that don’t have enough research grade observations for the AI to be adequately trained on them. If you want a change, then help make that change.

  3. The user is not required to accept any AI identification suggestion. Even if they do it’s not a bad thing if the first ID is incorrect as that places it into a category where someone who is interested in and knows that suite of species will see it and can provide a correction or confirmation.

There are a lot of complaints made about the AI identifications, but many of the complaints seem to come from a misunderstanding that the AI is meant to provide definitive identifications, which is simply not the case.

7 Likes

There is quite a gulf between what the ai “is meant to” do, and how it is actually being used, particularly by new users. For most fresh users with low observation counts the ai might as well be The definite answer.

1 Like
  1. Then on the first place there should be no sign “we’re pretty sure this is…” at all for any taxa.
  2. As I said many times above, for most gastropod taxa AI will still not work no matter how good it learning when it’s simply not possible without distribution data, unless it can see anatomy through the shell on photo.
1 Like

One would think all these problems could be alleviated if the ai provided higher level taxon IDs at whatever level there is actual confidence. As a side effect over time this would train users to expect less specific identifications.

2 Likes

Misunderstanding what that means and thinking it is a definitive answer is a user error, not one of the AI or iNat.

That said, I’ve proposed that AI IDs be marked with a flag saying “Preliminary Identification” or something similar and only count towards the ID count if they are confirmed.

1 Like