Facebook’s and Google’s Neural Networks Can Recognize and Create Images

Facebook and Google are both creating their own AI neural networks that can resemble the natural brain in many frightening ways. Their latest upgrade has already moved past the recognition of cars, faces, buildings and landscapes in digital photos.

These artificially-developed brains are starting to recognize spoken word, be able to make viable translations and even teach a robot how to screw a cap on top of a bottle! And in a twisted turn of events, they can now do more than just recognize images.

According to some new information released by Facebook on Friday, their neural networks can also automatically create little images of animals, cars and airplanes and have humans believe there are looking at a picture taken from reality in roughly 40 percent of the cases.

Rob Fergus, artificial intelligence researcher at Facebook, explained that the AI can differentiate an artificially-created image – such as abstract art – or an image that you would find on your camera; it is smart enough to understand the image structures and judge based on that.

Google’s initiative, however, seems to be approaching the AI matter from the opposite extreme, by teaching neural nets how to make real photos of real things into something fascinatingly surreal.

Google’s project was explained in a blog post, saying it all comes down to learning and enhancing familiar patterns in repeated way. Basically, if the program recognizes that a cloud looks even a little bit like a bird, it will enhance the characteristics of the bird inside the cloud.

It is fascinating to see how extremely detailed birds seem to appear out of nowhere, resulting in artificially-generated surreal art. It might sound like Google’s feedback loop’s only application is using it as a party trick, but it’s so much more.

Both Facebook’s and Google’s neural networks are working to advance the technology so they become even more alike human-like intelligence. Having them work with images is just a way of helping researchers visualize what the networks are learning to do, according to David Luan, director of Dextro, a computer vision company.

The images created by Google’s AI are also slightly unsettling, as the majority seems to be taken out of a drug trip gone wrong. The result of the feedback loop ends up cross-breeding birds with camels, or pigs with snails.

Researchers are rather excited by the larger point of these experiments, which is that the system is slowly but surely reaching the point of the so-called “unsupervised machine learning,” where machines can teach themselves without reserchers having to provide step-by-step guidance.
Image Source: Wired