Computer Vision should tell us how sure it is of its suggestions

I would be concerned that many users, especially new users, would misinterpret any displayed numeric value as some level of relative confidence. I concur with @kueda that some would overly rely on the *cryptic opinion of this black box" and not their own judgement. Student users would be one group of new users who might overly rely on displayed values.

4 Likes

Just to argue with myself, maybe including such numbers would allow people to make better choices among the black box outputs. Counter counter argument: the problem is over-confidence in a numerical rating when the right answer isn’t even on the list. Counter counter counter argument: that’s already a problem. Counter counter counter counter argument: we’re all doomed and should be spending our time learning post-apocalyptic survival skills.

18 Likes

On it: https://www.inaturalist.org/projects/search?q=edible

6 Likes

i think the scores are basically a measure of visual match and possibly some other factors like presence of observations of a taxon nearby. i think that’s why iNat staff don’t want people thinking of it as a probability or confidence.

so, for example, suppose you have 3 brothers A, B, and C. A & B are identical twins. suppose you take a picture of A and run it through a computer vision algorithm similar to iNaturalist’s. i would expect that CV to return scores that might be like this:
F: 0.97 – the family of A, B, & C
A: 0.95
B: 0.93
C: 0.65
D: 0.35 – D is the boy who lives next door.

so obviously, the CV couldn’t be both 95% sure the photo was of A and also 93% sure that it was of B, nor would it make sense to assign 95% probability of A and at the same time assign 93% probability of B, but by seeing the relative scores, you could see that the CV was saying that A and B were way better potential matches than C or D.

7 Likes

32 posts were split to a new topic: Chrome extension showing Computer Vision confidence

I like what @psium suggested. Perhaps it can be a % similarity [to other photos of which are RG]

e.g.
98% similar to [photos of] genus A
92% similar to [photos of] species X
91% similar to [photos of] species Y
75% similar to [photos of] species z
etc.

With an option to view the top ~10 ranking IDs in the Identotron on a separate tab, this will also show where the nearest observations have been to try and eliminate Australian species suggestions for African observations.

Identotron: https://www.inaturalist.org/observations/identotron?observation_id=49603263&taxon=51468#establishment_means=&order=&place=97392&taxon=51468

I agree, people would tend to read the higher “match” percentage the wrong way and jump on the identification without really looking closely at it.

This is already a bit of an issue (sometimes I even have to stop myself from doing it).

I think it’s best if the AI suggestions are left general. I like the “pretty sure” wording as well, that reminds people that it’s not a certain ID.

3 Likes

If you need some book references, hit me up…

3 Likes

I’m new to this community, so I won’t weight much here, but I think it might be a bad idea.
We have some Formicidae screenshot, they are a good example of why this is a bad idea.

Currently the CV is very bad at suggesting and identifying ants. I don’t recall the algorithm ever suggesting the right species as its favorite choice. I even had “we are pretty sure it is a [genus]” on an insect from a totally different order. Also almost every black ant are suggested as “Camponotus” Because they are all black, 6 leg 2 antennae 1 head one thorax one “wasp waist” and a big gaster. Most don’t have striking pattern like butterflies or birds can have for example.

Such a feature would give a false sentiment of confidence, as stated by other members. But it is already the case with “We are pretty sure it is”.

But worse, a red highlight would give the impression that it is unlikely the correct ID is actually among these “uncertain” suggestions.

Simply put : it’s a confusing system, especially for beginners.

4 Likes

yes. i think that’s exactly why showing the underlying scores would be enlightening – because the #1 choice in a list of bad choices is not the same as the #1 choice in a list of good choices. so if you can see the actual scores, you have better insight into whether you’re being presented good choices or if you’re being presented bad choices.

just for example, here are 3 of my own ant observations, along with the actual computer vision scores:

  • https://www.inaturalist.org/observations/43079992 (needs ID as Myrmicine Ants, possibly Bicolored Pennant Ant):

    we’re pretty sure it’s in this family:

    • (99.9) Ants (Formicidae)

    our top suggestions (combined score, vision score):

    • (40.2, 53.4) Dolicohoderus genus
    • (10.2, 4.5) Bicolored Pennant Ant (Tetramorium bicarinatum)
    • (5.6, 7.4) Acorn Ants (Temnothorax genus)
    • (4.6, 6.1) African Big-headed Ant (Pheidole megacephala)
    • (4.4, 1.9) Graceful Twig Ant (Pseudomymex gracilis)
    • (4.2, 1.8) Florida Carpenter Ant (Camponotus floridanus)
    • (3.2, 4.3) Arboreal Bicolored Slender Ant (Tetraponera rufonigra)
    • (3.2, 1.4) Crematogaster laeviuscula
  • https://www.inaturalist.org/observations/20826798 (research grade Red Imported Fire Ant / Solenopsis invicta)

    we’re pretty sure it’s in this family:

    • (100) Ants (Formicidae)

    our top suggestions (combined score, vision score):

    • (50.0, 30.5) Forelius genus
    • (8.3, 15.4) Asian Weaver Ant (Oecophylla smaragdina)
    • (6.0, 11.1) Mediterranean Acrobat Ant (Crematogaster scutellaris)
    • (5.7, 3.5) Argentine Ant (Linepithema humile)
    • (5.5, 10.1) Azteca genus
    • (3.5, 2.1) Crematogaster laeviuscula
    • (2.6, 4.9) Tropical Fire Ant (Solenopsis geminata)
    • (2.6, 4.8) Yellow Crazy Ant (Anopolepis gracilipes)
  • https://www.inaturalist.org/observations/44126816 (research grade as Eastern Black Carpenter Ant / Componotus pennsylvanicus):

    we’re pretty sure it’s in this genus:

    • (77.9) Carpenter Ants (Camponotus)

    our top suggestions (combined score, vision score):

    • (64.3, 61.3) Shimmering Golden Sugar Ant (C. sericeiventris)
    • (7.0, 1.8) Eastern Black Carpenter Ant (C. pennsylvanicus)
    • (6.9, 7.3) Giant Turtle Ant (Cephalotes atratus)
    • (4.0, 4.3) Eciton genus
    • (2.6, 2.8) Bullet Ant (Paraponera clavata)
    • (2.0, 2.2) Diacamma genus
    • (1.9, 2.0) Giant Forest Ant (Dinomyrmex gigas)
    • (1.8,1.9) Hairy Panther Ant (Neoponera villosa)

if using sessilefielder’s red-to-green gradient, remember that:

so most of the “top suggestions” above would be red to yellow, whereas the “we’re pretty sure” suggestion would be more green. hopefully in such cases, that would push most folks to select the green rather than the yellow or orange, if they were simply choosing blindly based on the system’s suggestions.

i also think if people could see that, say, bird suggestions tend to be very green, while, say, spider suggestions tend to be very red, then they would also be much more careful about relying on the computer vision for spiders.

or if they see two equally green birds suggestions, they might pause for a moment to consider why both are equally green before just blindly selecting the first choice.

of course, computer vision suggestions will never be perfect. there will always be mistakes, but i think showing the computer vision scores will help reduce (rather than increase) the likelihood that the community will adopt those mistakes.

9 Likes

I would have assumed that the scores are the classification scores in the final “softmax” layer of the neural network. In that case, every taxon would be given a score from 0.0 to 1.0, and all scores would sum to 1.0, like probabilities.

i think that’s hard in this case because of the hierarchical nature of the taxa. suppose your had an observation of a blue jay, and the existing algorithm assigned a score of .90 for blue jay, .95 for bird, and 1.0 for animal, what kinds of scores would your assumed implementation assign those taxa?

EDIT: nevermind – alex’s response below changes the way i have to look at things…

Hi folks,

Alex here, I’m one of the people who trains the computer vision system for iNat.

As tpollard mentions, the softmax function outputs in a format that is shaped like a probability. However, the output of the softmax function is strongly influenced by its input distribution. Given the imbalanced nature of the iNat dataset (we have a lot more images for some taxa than for others), these scores should absolutely not be interpreted as statistical probabilities.

I know tpollard didn’t suggest that they should be interpreted this way. I just wanted to make sure that the format similarity didn’t encourage someone to think about the scores in a way that isn’t warranted.

Thanks,
alex

13 Likes

If the CV thinks it’s 75% sure, the system should probably recommend identifying at a higher level. Guessing species IDs when not sure, is not helpful in my opinion. It just frequently leads to wrong IDs which often have to be overridden by multiple identifiers if the original ID is not corrected, which all too frequently happens.

As seen in cicadas, a more frequent scenario would be that the CV is 95% sure of the top pick, yet the user picks something else presumably because they are matching unimportant details (like color rather than certain pattern elements). If they see their pick has a 1% chance of being correct, they would probably be less likely to select it.

1 Like

It would improve things a lot if when the CV is pretty sure an ID is a particular species it would say so. Currently for the cicada Neotibicen superbus, which is easily identified, it just says it’s pretty sure its Neotibicen. See for example here: https://www.inaturalist.org/observations/94684158. Does the CV ever say its sure of a species? It definitely should in my opinion.

No it doesn’t, and I think that’s a good thing for two related reasons.

  1. The CV doesn’t know all species. It doesn’t even know all species with records on iNat! It only knows the species which had at least 100 observations when the model was last trained. This means that the AI’s calculated level of certainty may be totally inappropriate, if a picture matches only one species in the training data set, but would also match 10 or 100 similar species that didn’t make it in (e.g., they may be rarely observed or hard/impossible to ID from photos).

  2. “Pretty Sure” suggestions don’t take location into account at all. This means that these recommendations can be based on a match with a species that doesn’t live anywhere near you. Usually it’s pretty good for North America and Europe, but sometimes it’s still wrong. I can only imagine that places with relatively few observations, like South America and Africa, are much worse. I’m sure that leaving “Pretty Sure” at genus level greatly reduces the number of geographically inapproprate IDs on iNat.

Inappropriate CV-based IDs are a perennial problem here. They can flood CV-included taxa with clearly incorrect junk, and trigger-happy agreers can easily push them into Research Grade, where they get forwarded to GBIF’s database, and, critically, no longer show up to IDers by default. I understand your desire for “Pretty Sure” species suggestions, but I don’t think we’re close to being ready for that. I’d much rather have correct but vague IDs than precise but wrong ones.

4 Likes

What disappoints me is that - Seen Nearby - can mean - a single obs - with ONE ID!

That skews the distribution map, and cascades into - Seen Nearby, so mine is also that.

Computer vision is a suggestion. It is still up to us to evaluate and decide.

5 Likes

Old wrong ids are not helping that, I see how new observations get new wrong ids (and checking cv, yes, it shows species as seen nearby), when this “nearby” is actually three observations across whole USA, all in different states, none in the state where observation in question is. All are wrong, but seen nearby feature doesn’t help in using one far away obs as a sign of species being presented in an area (it can work that way with endemics too, or sp. found on one coast only).

3 Likes

That ‘Seen Nearby’ needs some quality control.
At least one or two Research Grade community consensus IDs before brightly suggesting ‘this is seen nearby’. Not! Actually.

4 Likes

100% agreement. ID suggestions based on “Seen nearby” should be more critical and require research grade observation or multiple observations with community ID. I think multiple community IDed observations is better because it is often difficult to reach a research grade (this is from my experience with insects, for different groups of organisms this will be likely different).
From my recent experience when I was cleaning some ID mess for insects which live only in northern America but were observed tens of times across the tropics of the world, these IDs were never even community ID - always just a single auto-suggested ID given by the observer. Hence, excluding observations without community IDs from “seen nearby” suggestions could be good enough.
@dianastuder @fffffffff this seems like a relatively easy thing to implement with large effect on the accuracy of IDs - shall we formulate it as a feature request? Or was it already formulated as a feature request?

EDIT: I actually though that “Seen nearby” is a parameter used for ID suggestions which does not seem to be the case. But I still agree it should be more restrictive about what is considered an evidence of nearby presence.

2 Likes