Add a "visually identifiable" scale parameter to set user expectations on confidence of computer vision IDs

AI suggests this is XY

We need to let new people know not to auto-agree
(do you know this is XY? Then click agree)

iNat is pretty sure - sounds definite and unambiguous. This IS what it is, no doubt about it!

4 Likes

Right - I’m not saying to prevent selecting it. I was really thinking along the lines of “setting expectations with the observer” to help them understand when a visual ID is actually quite difficult and unlikely unless they are an expert. I’m not a UI/UX person, but I’m confident appropriate steering/wording could be done.

I agree that this would be a challenge. One example of this I read recently that leaf miners, for example, are often easier to identify by the host plant and mining pattern than visually looking at the insect directly.

But, interestingly, it seems like the AI determines if it thinks an insect is a nymph or adult when making the suggestion. It would be good to surface those assumptions. like “iNat thinks this is a nymph Milkweed Beetle”. That would provide the users with more insight into the suggested ID and hopefully give them pause when it is clear that the observation is not a nymph, for example.

A possible solution for this would be to seed a tuple data structure, i.e. {taxonomic_level, life_stage}, and leaving unseeded tuples as it works today, and providing the extra insight only on seeded tuples. This gets a little more complex from a database perspective, so another idea would be to only seed in cases where ID challenge is consistent across all life stages - i.e. taking a minimally viable starting point from which you can then improve upon with future augmentations to the feature.

It doesn’t do that. It compares the photo to a set of photos which include larva etc and determines this looks like x.

It doesn’t say this is a larva, let’s go compare against all the larva photos I have access to and limit my analysis to those.

3 Likes

I agree that it would be hard to do with everything all at once. A good approach is to start simple. Maybe pick a few families, run an experiments, see if it can scale and be maintained. My view is maybe it never has to be completed. If it can simply be improved in some useful way, say, for some problematic taxa, that might still be a win.

My secondary thought is that with the idea of seeding, it ties in with the entire concept of machine learning. ML needs humans to seed the knowledge. By seeding, over time, ML would learn and could inform the AI which scenarios are not identifiable, filling in the gaps. It might take years, but that’s pretty much how it works.

1 Like