Both Google and Facebook have experimented with neural networks in order to generate art. Images are generated automatically depending on how the neural networks understand the objects.
Google explained that neural networks consist of layers of artificial neurons. When a picture runs through a network the first layer detects low-level content such as the edges of the picture. The following layer adds more information about the shapes which gives a better image of what is depicted there. What the final layers do is assemble the information from the other layers and do a complete interpretation of the photo.
Facebook creates rather realistic images. The company uses two networks in order to produce small thumbnail images. The first algorithm of the neural network generates pictures based on a random vector and the second algorithm checks the pictures for realistic images and does not take into account objects which look fake. So overall the pictures are very convincing. In fact they are so realistic that 40% of the pictures made people believe that they were real. After further refinements the pictures may become even more realistic in the future.
On the other hand, Google’s system tends to create abstract images. The researchers have named this technique Inceptionism. Google’s artificial neural network has a very advanced type of vision and is able to distinguish objects among one another. It can for example make the difference between a starfish and a banana.
Google generates art by allowing the neural network run wild and choose which visual elements it wants to emphasize. If the system analyzes a photo of the sky and believes that there are birds there it will amplify their traits until you will be able to see them. The images seem impressionist and surrealist.
The search giant calls these “dreams” created by the artificial neural network. According to Google this technique presents the way in which a neural network can carry out difficult tasks which require classification. It can also improve network architecture and see what the network has learned during the training.
Although the programs used by Facebook and Google to create art are quite skilled, they are not going to replace artists and photographers. However this technique may become a tool for artists because it offers a new mix of visual concepts.
Image Source: Google Research