Algorithm suggests unexpected taxon

Platform: Website

Browser: Chrome


Screenshots of what you are seeing:

Description of problem: The top suggestion is not in the indicated genus.

Without location information, the CV thinks your observation is most likely to be either Hemerocallis lilioasphodelus or Hemerocallis minor. The ‘we’re pretty sure’ recommendation does not consider location so it suggests Hemerocallis. You have ‘only show seen nearby’ checked. Hemerocallis minor has never really been observed in the United States at all, so it is not in seen nearby. Hemerocallis lilioasphodelus has been observed within 2 kilometers of the location of this observation in the same week of the year, so I am confused why it would not be listed in ‘seen nearby’.

1 Like

Those aren’t RG observations, which is what Seen Nearby uses.


I have never seen the suggestion algorithm contradict itself like this.

Here’s what it looks like if you turn off the “Seen Nearby” filter:

So the list above is ranked by visual similarity only. You’ll see it thinks the two Hermerocallis species are what it considers the most visually similar.

As @wildskyflower says, the “We’re pretty sure this is in the genus” suggestion is only using visual similarity, so I think it makes sense that it would show Hermerocallis there. If you turn onn “Seen Nearby Only” then Hermerocallis are removed from suggestions and U. grandiflora is the top suggestion because it’s the most visually similar and has been “Seen Nearby”.


Vermont has about 114 inat observations per square mile; out here in Montana where we have 1.8 observations per square mile so you get used to ‘seen nearby’ not being trustworthy.

Another relevant factor is that I think the CV is very confident it is Hermerocallis, but it is entirely possible it doesn’t even know about the correct Hermerocallis species; I’m no expert in daylillies but it doesn’t really look quite like either of the species suggestions to me.

For me, a similar inat observation that the computer was ‘pretty sure’ was a specific genus that had nothing seen nearby turned out to be a state first occurrence record of an invasive species not previously documented anywhere within hundreds of miles of the state. A specimen from a construction site near me was accessioned to the state herbarium and published as a collection of note. I’m glad the computer vision was ‘pretty sure’ and that I was confused by the situation enough to tag an expert.


This often happens, if this is a bug I support it.

It seems to me that there’s a UI design problem here where the “We’re pretty sure” suggestion ignores the “only show seen nearby” preference, therefore leading to an apparent contradiction.

It could be fixed by making the “We’re pretty sure” suggestion also honor the “only show seen nearby” preference. But it could also be fixed by leaving the behavior as-is, but beneath the “We’re pretty sure” suggestion include a disclaimer like “not yet seen nearby”.

1 Like

It looks like this, but ‘this’ is not seen here - is a definite conflict of common sense.
Ploughing thru the South African Pre-Mavericks to sort out, no, not Siberian, or Californian or …
For newbies - pretty sure - sounds like this is the right choice. Click. :slightly_smiling_face:

Would “not yet recorded nearby” work better? Or do you think the “We’re pretty sure” choice should not be shown if the taxa hasn’t been RG-observed nearby?

I’m not American and at first I thought - pretty sure - was emphatic. Definitely this!

But I am coming round to - it is meant to express a degree of uncertainty? But then that is getting lost in translation. And English is my first language.

I would like an obvious gap between - seen nearby, looks right, suggest this

no idea but this Siberian looks the same. It could be invasive or cultivated, but 9 times out of 10 - not a useful suggestion. And if someone agrees we need 5 IDs to convince the CID algorithm. That is an exercise in futility for some taxons.

“Pretty sure” definitely expresses a degree of uncertainty, but not a very big degree. Perhaps making that language weaker in the case of a suggestion for which there are no nearby RG observations would also be a significant improvement.

I wonder if it would be possible to get any statistics on how often the “pretty sure” suggestion turns out to be right/wrong in the case for which there are no nearby RG observations? Of course, defining “turns out to be right/wrong” in this case isn’t straightforward either.

For my understanding - the intended meaning is

It could be This taxon

Which would make the UNcertainty obvious.
On a scale of 1 to 10, I would rate ‘could’ as a 6, but ‘pretty sure’ is an 8 or 9. Where’s the problem? Lost in translation.
Where is your ‘pretty sure’ on 1 to 10?

For me, I think of “pretty sure” as a 7 or 8. “Quite sure” would be a 9, and “sure” would be a 10. “Could be” comes across as extremely vague to me, anywhere from 1 - 9.

I thought pretty sure is between 94 and 92%. You can see it in the diagram as the upper pink line marked with CA (Common Acestor) and is based on Computer Vision only. No nearby observations are used. (I see the data might be a bit old, based on Computer Vision model in 2017)


I do lean heavily on the app - which shows me, across a green to red scale, how sure the CV is. And my own knowledge - no, that is definitely not a seal halfway up the mountain!

App was built for us by an iNatter, and is updated with extra features as and when.
Recently - copy geocoordinates.

1 Like

If you want to translate ‘‘We’re pretty sure it is in the genus’’ is important to know what exactly is mentioned with this line… I found out yesterday…

1 Like