I remember hearing about 6 months ago that a new training run of AI was imminent. I tried to use it today on some cicada species that should have been over the threshhold of 100 observations, but AI still doesn’t know about them. At least one species was previously way below the threshold and wouldn’t have been covered in previous runs. Has the AI not be retrained in a long time? What’s the status on training?
I have also been eagerly awaiting this–we’ve made a huge amount of progress getting galls over the 100 obs threshold this year and I’m looking forward to using the new CV on this year’s observations.
(I changed your title because while we (including staff) sometimes use “AI”, “computer vision” and “machine learning” are more accurate.)
Training was delayed last year due to the pandemic, but it’s currently underway now. Might be another few months until a model is ready, though.
Does this mean the cutoff for observations/taxa to be included in the new model has already passed?
The data will be used anyway as the model uses species insertion…Other projects are very busy with morfologie (gal, imago, larve) but in iNaturalist this morfologie seems rather limited. Is the type not important for galls ? Somewhere they (Loari?) wrote that the photos from end 2019 were used… The other project was using a training machine with 10 videocards…Does iNaturalist also has such a super VideoPC ? Or is the training just on one (Nvidia) videocard ? https://forum.inaturalist.org/t/carbon-footprint-of-inaturalist/11126/12
In another computervision model the admins were busy telling/tagging every photo if there is an imago or larve on the photo which could be used by the training model. So i thought it might be neccassry to tell the model that it is a photo with a Gall
Not sure what you mean by type?
Yep, we created the training set for the new model in September. That’s the cutoff for inclusion.
When is the next cutoff?
We don’t have the next training run scheduled - there is too much uncertainty right now to decide that: how long until the current training cycle ends, when we can get back into the office to build our new server for future training runs, how our current production and development training runs compare and if the new dev software is ready to take over as the new production training system or if it needs more work, etc.
Vaguely, I can say “sometime after the current training run finishes.”
I think the different types are gall vs adult vs larva, etc. I think the model already handles this as far as species identification, but the CV could also tell you that it’s a gall in addition to telling you the species. For that to be done, it would have to be labelled (e.g. “gall”) and that be made part of the training model. I doubt it is.