AI Gone Wrong: the Great British Bake Off Becomes an Optical Nightmare
What she saw instead were gruesome images that are more fitting for a horror show. Nothing charming, sweet, or delectable about these photos.
What went so wrong?
Researcher Janelle Shane shared her experience on her website, AI weirdness, on which she explained how she trained NVIDIA's StyleGan 2 system on photos of the famed show's bakers, their baked goods, the tents, trees in the backdrop and even a random squirrel.
Instead of generating realistic replicas of faces the AI system instead hashed out "portraits" that are worthy of a vivid horror show. The words grotesque, mishaped, and twisted come to mind in this instance. In fact, they're so bad that they're good.
According to Shane this is a perfect example of what you can and can't do with deep learning technology.
It wasn't for lack of data, as Shane had trained the system with 55,000 images taken from the show. The issue was when faces were introduced into the project that the system hadn't memorized. So the images were not always perfectly centered, they were different sizes, there were hands at times, a loaf of bread here and there, a tent, or a squirrel. All things that the system did not recognize.
So rather than creating faces the system erased them. Talk about system overdrive.
Shane didn't want to give up there though, so she kept training the system, which unfortunately did not bode any better. Shane wrote about the issue on her site, "This is the usual outcome when you train a neural network for a long time -- not an acceleration of progress but a gradual stagnation."
"The baking show images were too varied for the neural net, and that's why its progress stopped, even with lots of training data."
By looking at the AI-inspired images it's hard to tell what's what. If anything, it can become a fun guessing game, but that's about it, unfortunately.