🔥 Hot or Not? Beauty in the Eye of the Mainstream
I know it sounds a bit callous, but hear me out. After I made my first three little web apps using fast.ai neural network methods, for classifying cats vs dogs, 3 different types of bears, and 3 human emotions/facial expressions, I watched my children spend literally HOURS playing with them. They put EVERYTHING they could find as input to see what the model would output. This was actually a very illuminating look at what the model sees.
They were putting in all sorts of anime characters and random objects, and I started to see a pattern for each of the apps / models. And this made me realize how through some experimentation, I might be able to create more accurate modeling for human faces. They are, in fact, one of the hardest to classify.
So I took 400 images that the almighty interwebs deemed "attractive", "average", and "unattractive". I then created various transforms on each image as a means of data augmentation. I wanted to give the model as much data as I could without needing millions of labeled photos, which by the way are really hard to find, at least for this particular goal. I guess no one wants to be the one officially labeling beauty. I get it.
And while I felt a little guilty at first even making this app, knowing that it could be easily misunderstood, I realized some really amazing things. The computer can see genuine facial expressions versus forced. It is also incredibly senstive to angles even though I used so many transforms of the images. And I know there is much more that I will find as I work with this data. It has, after all, been less than a day since I finished the project.
I look forward to discerning more, and I will update you all when I do. But for now, this is just a little glimpse!
And honestly, I was rather offended when my app did not recognize the unbelievable beauty of my standard poodle, Ragnarök. 🤣
