What a painting! Is it Rembrandt, Rubens or AI…

Reading Time: 3 minutes

The story about GANs (Generative adversarial networks) creating paintings worth $432,500.

At the end of October the famous auction house Christie’s in New York sold a painting made by GANs for a surprising amount of money – $432,500. The average person may not know what generative adversarial networks really means. For this reason I will start with a short explanation of this class of artificial intelligence.

Let’s imagine two algorithms playing a game with each other. One of them uses machine learning to gain a knowledge about appearance of a certain object. In practice this algorithm may be given a 1000 photos of different kinds of apples and after that it should recognise shared features (shank, pip, encountered colours etc.). With this knowledge it starts to create new pictures representing the same object. During that the second algorithm (which also saw photos of apples) sees the works of the first algorithm and it is telling if objects in this drawings look similar to an apple or not.

Sounds simply? Let’s make it a bit more complicated! The first algorithm is called generator, the second is called discriminator. Both of them are neural networks. The generator creates random synthetic output (it can be an apple, face or an image) while the discriminator makes distinctions between green apple and the others or between painting representing abstract art and the others. After such interaction both algorithms are enhancing their skills.

https://poloclub.github.io/ganlab/

https://poloclub.github.io/ganlab/

In the picture we can see green line made of dots. It is a real sample. The second picture shows how those two algorithms read given data. The generator creates gradients while the discriminator makes predictions of sample. As the result we see a new line, which length do not vary in length and angle of incidence of the real sample. However an observer can easily tell, that two lines in two pictures are not the same, but have some common features.

Previously mentioned painting “Portrait of Edmond Belamy” which was sold at Christie’s was created by GENs. Creators of this algorithm (Hugo Caselles-Dupré, Pierre Fautrel and Gauthier Vernier) fed the system with a data of paintings from 14th century to the 20th. The GENs accumulate data of around 15 000 portraits and using the technic of “two neural networks playing a game” created a completely new portrait.

https://www.telegraph.co.uk/news/2018/08/22/computer-created-portrait-sold-christies-sale-marks-arrival/

 

Such innovations generate new questions. Is it possible to create a piece of art without an artist? Where is the border between creating an art with AI and copying existing paintings? Who is the creator – GENs, authors of all paintings used by an algorithm or the developer of GENs?

In my opinion one of the most important is “What are the others applications of GANs?”. Nowadays GANs can be helpful to visualise different possibilities of designing shoes, clothes, interior designs or scenarios of computer games. It helps also to improve blurry images, such as old photos or destroyed paintings. Probably it is also capable to create things, which we can not imagine yet. It is a tool which can influence our next century.

https://www.technologyreview.com/s/612501/inside-the-world-of-ai-that-forges-beautiful-art-and-terrifying-deepfakes/

https://www.christies.com/features/A-collaboration-between-two-artists-one-human-one-a-machine-9332-1.aspx

https://gandissect.csail.mit.edu/

https://www.youtube.com/watch?time_continue=156&v=yVCgUYe4JTM

https://poloclub.github.io/ganlab/

2 thoughts on “What a painting! Is it Rembrandt, Rubens or AI…

  1. Willoński Artur says:

    Thank you, for reminding this aspect of AI. We are all thinking about Artificial intelligence in a way of building robots or business predictions, but that is not all. Now I am wondering what music would be like, while made by AI.
    That might be interesting, beautiful or simply terrible, all depends on Data you use 🙂

  2. Piotrowska Maria says:

    I have the same strong curiosity. For me the most interesting thing is that it might be possible to predict who would like a certain song before even creating it. Also I find interesting that based on huge database about music taste (such as Spotify) you could produce personalised songs for all users around the world, everybody would have their own songs – different from the others.

Leave a Reply

WordPress › Error

There has been a critical error on this website.

Learn more about debugging in WordPress.