Remember last week when we talked about the strange images created by Google's neural networks? Well, since then, the authors of the original article received all sorts of questions, not only from computer experts but also from artists wanting to know more precise details about how each image was generated. The Mountain View giant decided to respond by releasing part of the code under the DeepDream project, so that everyone interested can investigate it.
An artificial intelligence making art. Multiple layers that not only interpret data, but see things where there seems to be nothing. Buildings in the sky, animal heads, leaves transformed into birds. The world of surrealism is probably rubbing its hands together at these results. If there is something we must concede to technology and computing in general, it is that it has served as a bridge to connect completely different genres. How else would it have been possible, for example, to transform into an artistic initiative the idea of storing digital music in DNA? A good part of the artistic universe began to pay more attention to computers when it saw Andy Warhol on a Commodore Amiga. Another example: a small child sits in front of the computer. Which program is most likely to be opened? Paint. Definitely there is something between art and computing. The images made by Google's neural network only accentuated that relationship.
DeepDream Code Released on GitHub
The questions are inevitable. How? When? Why did it choose that? Google Research had no better idea than to publish the code on GitHub under the name DeepDream. The code appears to be an IPython notebook and is based on the Caffe deep learning framework developed by the Center for Learning and Vision at UC Berkeley. In essence, everyone interested can take this code to generate images inspired by the neural network's results. Once the environment is configured (the installation process is relatively complex, to be honest), the user-artist can provide an image, choose the layers of the network to optimize, the number of iterations and the zoom. Alternatively, DeepDream allows connecting networks that have already been pre-trained.
The only thing the code's authors ask is that images be shared on social networks under the hashtag #deepdream, so that other researchers can see the results. If the final product is as good as it promises, I figure it's only a matter of time before someone offers paintings and prints on the Web...
Official announcement: http://googleresearch.blogspot.com.ar/2015/07/deepdream-code-example-for-visualizing.html
DeepDream on GitHub: https://github.com/google/deepdream
