A lot has been said about this all over the internet. Images, where the labeling program sees too many eyes in the sky, too many dogs in a horse and so on. Quite disgusting, I concur!
On top of that, this machine labeling is already more accurate than some humans at recognizing what’s in the picture. Another blow against sanity?
Not really. In fact, it is just a poor early approach. Overfitting is everywhere. A better version of the same algorithm should learn that it is unlikely for a horse to have more than one eye on each side. And that no doberman dog is in the sky hovering over the landscape. Unless it is quite a distinctive dog, and not a fractalized weasel monster.
Some old OCR software may have had the same problem, reading something as “TO7AL SUM”. It is unlikely that there is “7” there – much more likely it is the letter “T”. Newer, better OCR products just don’t make this mistake anymore.
Context is important, the rules of the context are very important. They should be learned and used. Then, this stupidity or insanity will go away.
When every small square of almost any picture will be computer-labeled exactly, like horse nostril hairs or reflection of the house in the horse’s left eye and so on … THAT will be a fine, sane, not crazy software package.
As Patrick Wilson once said: I want this picture to be automatically labeled as toasting and this one as drinking! A small but essential difference we understand quite well.
Then, it will be the time to brag!