AI Reimagined the Great British Bakeoff as a Weird Acid

In her latest experiment, AI researcher Janelle Shane trained a neural network to use 55,000 screenshots from the show featuring images of bakers, pastries, tents and even random wildlife. The neural net spit back what it thought best represented the quaint baking show, and the results aren't tasty.

Shane used a state-of-the-art image-generating neural net called StyleGAN2 that's rather good at understanding the concept of human faces. But it doesn't do so well when you add in human bodies, cakes, pastries and tents.

The first thing the neural net did after it was fed all the images was erase the human faces from The Great British Bakeoff screenshots. Apparently, the neural net was not only confused by human faces that weren't positioned dead center in the screenshots, it was also having difficulty figuring out external shapes of baked goods and the interior of the tent.

Neural nets are great at understanding patterns, so when presented with an image of one pie or one human body, it needs to replicate it over and over in the same screenshot to show off a pattern. That means human bodies in a screenshot might end up with extra arms.

"A neural net usually builds images by stacking lots of repeating features on top of one another, fine-tuning the balance between them to produce objects and textures," Shane wrote in her blog. "If it gets the balance slightly wrong, individual repeating features tend to pop out."

The most amusing part of this experiment was seeing what the neural net thought was the ideal baked good. Some of the more memorable items include a cake with weird holes, hoverbread and a blueberry pie with way too many layers. Yum!

Shane's previous neural net food experiments have proven to be both bizarre and entertaining. Shane trained her neural network to come up with weird Harry Potter pie creations, unusual cookie names and Valentine's Day candy heart sayings.