How to lose confidence in AI, or how I bullied with Gradient Photo Editor



The article is more of a sketch and leaves room for experimentation.

Hello everyone. This weekend, I saw the Gradient Photo Editor application, namely a function that allows you to evaluate your DNA using the latest AI technology from photography! An extremely accurate algorithm will analyze facial features and determine ethnic origin (description taken from offsite ). I, Mlista, could not help but be intrigued, and of course I downloaded the application and started messing with it. After the pictures of myself beloved ended in the course went masterpieces of world painting. And then surprises awaited me :)

Who cares - welcome to cat.

The idea with Picasso’s portraits of women failed right away, so I took the path of least resistance - I took Mona Lisa Leonardo da Vinci, Frida Kahlo of the artist of the same name and a girl with a pearl earring from Jan Vermeer.

For the purity of the experiment, before uploading the photos to the application, I made a forecast for myself which countries I expect to see.

I suggest that we first make a prediction, and only then find out the result.

Find out the ethnic origin of the heroines of the paintings


I was delighted and puzzled at the same time. Still, in each of the heroines there were not just Slavic roots, but, in fact, native, Belarusian! And this despite the fact that by the time of writing the paintings we were still far from self-determination as a nation ...

Over time, the euphoria passed, but the feeling of doubt remained. I, like you, noted that I didn’t draw a grid on the face for a girl with jewelry. It's hard to say why, but! this leaves room for all sorts of pranks: why not drive this picture one more time? Take a screenshot and one more time. And yet ... Just look at the reproducibility of the result :)

The question of reproducibility of the results for me personally is controversial. You can notice that from the first to the ninth screen, the skin tone changes - it becomes darker. I’m not sure what this is connected with, naively believing that the screenshot just takes a screenshot, saving the image in its original form. One way or another, it is worth noting that I began to feel the catch - with the increasing tan of the girl, the Scandinavian roots are becoming more apparent ... And this goes against the way my northern friends look. Well, the diversity of countries cannot but alarm.
Since this reproducibility test failed, I decided to try differently. This time, we’ll download several different Frida’s self-portraits with a thorn necklace and hummingbird and drive them through the algorithm.

My naive soul believed that there could not be such a cardinal difference, but the fact is that it does exist.

In addition, I uploaded photos of slightly promoted athletes, actors, scientists and politicians with enthusiasm discovering what a talented nation I was fortunate enough to be born. What is surprising is that my friends from Russia, Germany, Sweden, China and the USA see different results ... Probably, all this is the machinations of the Masons.

Conclusion


In conclusion, I would like to note that IMHO is not and cannot be exceptionally accurate as stated on the (description taken from the offsite ) algorithm. Perhaps a small dataset, perhaps limiting the development of ML or some other reason did not allow to solve this problem as intended. However, it should be understood that not everyone understands what ML can and what not. And if you play on people's fantasies about what AI has already achieved and can - then in the end they can be very disappointed in its real capabilities. And then the customer comes and asks to realize the completely unimaginable, because in his opinion so much has been done in this area.

Be careful not to sweeten the pill with geostated data.

PS In the comments sent - could not share.

They wrote that the floor matched :)

All Articles