Deep learning is a type of machine-learning algorithm based on non-linear, networked representations. Its latest application has enabled computers to make some highly unsettling artistic creations.
For over two months now, images from web giant Google’s DeepDream programme have been setting the Internet alight. DeepDream, featuring standard landscapes and well-known pictures bristling with psychedelic forms (link in French), hybrid dreamlike animals, multi-coloured arabesques and other fantasmagorical apparitions, which almost make the viewer feel as though s/he has been taking psychotropic substances, has really intrigued online visitors. However, this new tool that everyone is talking about was not fundamentally created with art in mind. It is a deep learning project (link in French) under which a programme that can recognise, identify and enhance patterns in existing images has been developed. Proceeding by association, the software extrapolates one form from another. In other words, if a motif shows a vague resemblance to the head of an owl, the programme will make a whole owl appear, creating disturbing, other-worldly dream-like pictures. A team of German researchers at the University of Tubingen has just has just completed a similarly intriguing experiment, also based on machine learning. The result of their work is an algorithm capable of creating out of simple landscapes pictures that could be mistaken for works by great masters such Van Gogh, Munch and Turner. In order to do this, the researchers fed the computer with dozens of pictures by different painters. The software digested and analysed the pictures so that it could recognise the style, colours, approach and brushstroke technique of each great painter. Subsequently, when fed any landscape image, the software is able to retain the image content – the main elements and their spatial positioning in the picture – while copying the style of one of the artists. The results are quite amazing.
Learning by storing multiple representations
So what exactly is this ‘deep learning’, which is capable of creating such spectacular renderings? The method is basically a supervised learning exercise for the algorithm: if you want a programme to learn to recognise a cat every time it sees one in a picture, you feed the software a huge number of images of cats, so that it will pick up on the various characteristics of the animal and then be capable of recognising one in an image that it has never seen before. A special feature of deep learning is that the computer proceeds layer by layer, i.e. it cuts up the image into various slices, working at each successive stage with ever-greater precision. (Yann Ollivier, a researcher in artificial intelligence at the French National Centre for Scientific Research, explains this in detail (in French) for those who wish to know more.
Food for thought about artistic, and other types of, original creation
The deep learning approach can be used for many purposes. Google Maps uses it to decipher text found in landscapes and cityscapes, such as street names, while astronomical researchers use it to recognise galaxies. Applied to artistic purposes, deep learning can actually be said to resemble the creative processes in the human brain, which certainly gives us something to think about in the context of art and our conception of it. Other surprising ventures on which French newspaper Le Monde has reported (in French) recently include a programme capable of writing its own musical compositions…in the style of Antonio Vivaldi!; another programme which can improvise jazz pieces; a third that composes short stories; and last but not least one which concocts cooking recipes. Artificial intelligence really does appear to be getting very close to human intelligence.