What Neural Networks "See" - News

Google's DeepDream service enables image processing through a program that "sees" strange animals and objects in a picture.

As Wikipedia writes, "the program was written to find faces and other patterns in pictures in order to automatically classify them." But if the program continues to set on the picture for a deeper search, then it begins to find faces and patterns, where they, in fact, do not exist. Well, or for lovers of philosophy, where a person does not see them. Or maybe the avocado really looks like this?

In this case, the pictures are obtained similar to the visual hallucinations of a person under psychedelics (according to experienced users). This fact is curious in that it can mean the functional similarity between the human visual cortex and the artificial intelligence neural network.

This is what the original avocado image looked like, and above, you can see what I was able to do with this image using DeepDream (level 7, very deep).

In another recent study, scientists (Radford, Metz, & Chintala, 2015), by training a special type of neural network, generative adversarial networks, have taught a computer to generate designs for rooms (quite good, by the way), and people's faces.

So, giving photographs of smiling female faces, they were given the task to highlight the concept of "smile" and connect it with the concept of "man". The result is a collection of smiling male faces of nonexistent people. Not perfect yet, of course, but networks learn quickly.

As the author of the article in New Scientist (Aron, 2015) notes, very soon we could search for pictures in Google by description - the picture itself will be generated on the fly and will be absolutely unique.

The day is not far off when artificial intelligence will create much better paintings and designs, both buildings, interiors, and any mechanisms and devices, and everything else. What are we going to do - this is the question that is time to start answering.

Literature:

• Aron, J. (2015). Computer's imagination creates human faces. New Scientist USA. December 5, 2015. Radford, A., Metz, L., & Chintala, S. (2015).
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. arXiv.org.