Google and Facebook have taken out their dispute over the reins of the Internet in pretty much every possible aspect, but their latest endeavors might surprise even the most astute analysts – the two companies are now racing to complete artificial intelligences capable of dreaming and even painting.
The two Silicon Valley companies are both in the process of creating neural networks simulating parts of a human brain, which can for example identify objects from digital images, recognize speech and translate language and are even have teaching capacities. However, both companies recently announced that their machines also boast creative capacities.
In Google’s case, a blog post earlier this week signed by Software engineers Alexander Mordvintsev, Christopher Olah and Mike Tyka detailed the process. Apparently, Google are using a technique they call “Inceptionism” to make the A.I. recognize objects which might be similar to others. If two objects look similar one to another, say a bird and a cloud, the A.I. will scan the image multiple times and make it look more like the other until it reaches a detailed version.
But things get interesting when the machine does not have a set object/feature it should amplify, and is let to enhance what it recognizes on its own. This would have the image become more and more abstract with every scan made of it, and the results shown in the blog are very similar to paintings.
„The results are intriguing—even a relatively simple neural network can be used to over-interpret an image, just like as children we enjoyed watching clouds and interpreting the random shapes” is noted in the blog post.
Facebook on the other side has not gone as grand as Google’s artificial intelligence paintings, but its multiple A.I’s can create 64×64 pixel images of basic objects such as animals, airplanes or cars. However, they are reportedly realistic enough to deceive about 40 percent of respondents when they are asked to judge their provenience.
One interesting part of the Facebook experiment is a training system developed by the company which sees the neural networks trying to fool each other in turns, with one trying to present a realistic image with the other judging whether it is of human or A.I. provenience. This is described as a competitive type of training which, unlike Google’s more artistic approach, ends up with incredibly realistic pictures created by the A.I.
Image Source: Google Research