Recommendations on improving the AI algorithm?

[zoology123]

Long but this is just responses to some things.

@paul_dennehy

I don’t think there’s any way to manually “tell” the CV to change its suggestions. It’s just a learning model that matches unknown images with a dataset of known images. As far as I know, no one can change what the CV spits out on a case-by-case basis like this. The only solution is to make corrections to the photos that have been misidentified thus far, and wait until it learns. It can be a long, tedious process, but it’s doable.

Everything done has to be indirect. So in a way yes you can manually tell it to change just in an indirect way. I alone got the CV to forget the taxon Diamesa for example. Also if you know what you are doing and set up some plans, you can go as far as to actually predict and plan what taxa gets learned. I have been predicting and planning which Chironomid taxa gets learned for over 6 months now. Other than one quirk I don’t understand yet, it’s actually been pretty easy to predict when a taxa may be learned. Just take time and effort.

@thomaseverest

Yes, by adding correct IDs as others have said. The CV is updated every month or so, so things should continue to improve.

This is a the endgoal, but there are so many other things one can actually do to get there.

@calvertm

The model takes into account not only appearance but also physical location

They are separate models that work together, what you are referring to is the Geomodel. It has some quirks just like the CV.

This also means one has to think about training both the Geomodel and the CV itself to get good learning results from the CV.

@AdamWargon

What DOESN’T happen, AFAIK, is the CV analyzing WHY they opted out, and looking at their unusually detailed and specific observation comments about DNA, microscopy, etc., and then using that information to improve its own accuracy.

The CV in a simple definition is simply an image matcher. It takes no details like that in.

@spiphany

biggest sources of problems with the CV is the fact that it is not trained on higher-level taxa if a lower-level one in that taxon is included.

This is in my opinion one of if not the biggest weakness and flaw of the system. It forgets higher taxa if leaf taxa are learned. It requires some quite silly planning to get around the negative effects of this.

@ShotShot

Anyway, I think suggestions for better onboarding (through nudges and popups and whatnot) were briefly considered in the distant past, but never touched upon again ever since. And I fear that anything turning users away from CV use is a big no-no, considering iNat is built around it.

Onboarding can only solve so much. As an analogy. Imagine a street with a speeding problem. Everybody should’ve gone to driving school and got a license, they should know not to speed. But some people still do it. If you focus on the design of the road itself though rather than education you can actually get much better results, Make it narrow, perhaps windy, with speed bumps. People will actually go the speed limit. INaturalist Should design the CV to better accommodate peoples actual behaviors. Education can only go so far and has to be constantly done to each new user.

@vbjanos

It ignores taxa that do not have enough RG observations to be included.

RG is not needed for taxa to be become eligible to be learned.

@vbjanos

Why not make CV more restrained

The CV needs community controlled restraint. Where taxa that have issues can be addressed by curators and the main identifiers bringing up these issues. A blanket restraint would be terrible.

@eyekosaeder

While I get your point here, I don’t think CV is completely unbiased. It will naturally pick up biases because the data it is trained on is based on humans with biases (if I understood how CV works correctly, at least). Unlike a human it’s consistent with the biases it has, though.

The CV can have quite a lot of bias. If a single observer with a single camera made all the observations to train a taxa. That taxa would be trained on just their photos. This means that photo quality, technique, settings, lighting, can all have an effect.

@someplant

limit the CV

Strongly agree with the implementation of a system that allows the community to discuss and implement locks on certain taxa. This alone could be an extreme improvement for taxa that are out of control with misidentifications or just taxa that really shouldn’t be in the model. Like Covid-19, or things that need microscope, etc.

@cyanfox

Definitely allow hybrids to be considered again. Excluding hybrids is really just a recipe for hybrids being misidentified as one of their parents.

If hybrids are leaf taxa of a parent species. Their inclusion will cause the CV to unlearn the species. The CV only learns leaf taxa.