Monday, 30 September 2024

AI - how well does it work?



I wonder how many of us are using the AI facility on iPhones - the little bug which comes up when you click the 'i' information icon below an iPhoto.  I've come to it only recently because of being slow to update my phone, but it seems trustworthy on obvious insects such as the Brimstone butterfly and the Brimstone moth.

As a little exercise, I'm posting a series of small, brown moths - the bane of my life - from a recent night's trapping with their AI description. Over to you Dave, and other experts to mark them as correct or not. In the top composite, looking from left to right in successive rows, the phone says: Willow Beauty, Willow Beauty, Lunar Underwing, Straw Underwing/Clouded Drab (though judging by Phil T's recent post, I think this may actually be a Deep Brown Dart), Double Square-spot, Archips podana,  White-point, Square-spot Rustic and another Lunar Underwing.

For the second composite below, it says: Pale Mottled Willow, Lesser Yellow Underwing, Square-spot Rustic and a third Lunar Underwing.

Im really interested to know how well it's done and mightily relieved if it's as generally reliable as appears to be the case.  Martin Wainwright, Thrupp, Oxon



4 comments:

  1. Hello Martin,

    There are two errors in your first collage, Deep-brown Dart as you've suggested and the one next to it in the same row which is Autumnal Rustic and not Double Square-spot.

    I've not tried using the method you've highlighted here, but I have had very limited experience of the apps 'Seek' and 'Obsidentify' (I'm sure there must be others too) which are aimed at providing IDs for animals/plants. I found that 'Seek' from iNaturalist made far too many errors when it came to moth identification, but 'Obsidentify' does seem to have potential even though it could still make really silly mistakes (for example, an image of Mediterranean Brocade it identified as a Sand Lizard!). I imagine that they'll all get better with time and, even now, could be a useful aid especially for beginners. The important thing is to recognise their limitations and not rely on the ID given as being gospel, but use it to maybe narrow down the field.

    ReplyDelete
  2. Many thanks as ever, Dave, and that's an object lesson. I've known the Autumnal Rustic for years and it's highly distinctive, so you've hlpfully highlighted the great danger around all AI - that we may get lazy. It's going to help me to narrow things down though, I think, in the same way as Flying Tonight and others' comments on this blog. All warmest and thanks again, Martin.

    ReplyDelete
  3. Hi Martin,
    You have nicely illustrated the potential problems with photo recognition software. Recently there has been some discussion of one of these apps (Obsidentify) among experienced Berkshire moth recorders. The conclusions were very much as Dave has already noted, but it might be useful to note the following:
    (1) The app's ability to identify a species will of course depend on the quality of the photo, although it will give an answer even if the photo is poor. To the unwary this clearly leads to potential mis-identifications. I have found that the amount of background included can alter the percent identity, or even alter the identification, so it is also important to crop the photo to include only the moth.
    (2) Some species are probably not in the database used by the app e.g., the adventive Tachystola acroxantha from Australia is common is some parts of our area, but no matter how good the photo it will not be recognized (usually identified as Carcina quercana). This is probably because the app originates in Europe (from a non-profit based in the Netherlands) not the UK. Further examples are given in this link from 2022:
    https://notquitescilly2.blogspot.com/2022/11/obsidentify-truth.html (thanks to Mark Calway for these examples).

    ReplyDelete
  4. Thanks very much John, that's really interesting. I'd cottoned on to the cropping after some rather weird guesses from the iPhone robot, though none as wild as a lizard! I guess that they will gradually improve as more and more data is fed into them. I'm going to take a screenshot now and see if it can spot a Porter's Rustic All warm wishes Martin

    ReplyDelete

Note: only a member of this blog may post a comment.