We are figuring out whether cars will oust living artists from the market.
The organizers of the Christie's auction in New York planned to sell the Portrait of Edmond Belamy for $ 10,000. However, the result exceeded all expectations. The starting price was exceeded many times and the painting went under the hammer for $ 432,500. What is the name of the author of the painting? What is this rising star of world painting? Unfortunately, it is difficult to pronounce the author's name for a person without mathematical education. Since the creator of the canvas is an algorithm that is expressed in a complex algebraic formula (see illustration). It is this formula that flaunts in the corner of the portrait, where artists usually leave their initials.
This formula flaunts in the corner of the portrait, where artists usually leave their initials.
The algorithm got carried away with abstract painting
“We trained the algorithm by loading data from 15,000 portraits written in the period from the 14th to the 20th century,” says Hugo Casell-Dupré, a machine learning specialist, one of the project participants. - When creating the portrait, a generative adversarial network was used (from the English Generative adversarial network, abbreviated as GAN). Its principle of operation is based on the competition of two neural networks (for more details, see COMPETENTLY).
Frankly speaking, the masterpiece looks, to put it mildly, a bit unfinished. But the “godfathers” have their own explanation for this. In their opinion, the algorithm was carried away by abstract painting, so the portrait resemblance was not particularly worried about it. French programmers believe that by analyzing the work of artists over 7 centuries, neural networks realized that art progresses along a well-defined trajectory - from form to abstraction. And they decided that it was necessary to correspond to the trends of the times. Moreover, in abstract painting there is more space for self-expression and novelty.
Promotional video:
The same Edmond Belamy …
The question arises: who is Edmond Belami, who received the attention of artificial intelligence? In fact, this is a fictional character named after the creator of the GAN programmer Ian Goodfellow from Google. His surname is translated from English as "good friend". The phrase "belle ami" sounds exactly the same in translation from French. However, the developers have at their disposal not only a portrait of Edmond Belamy, but also his numerous "family". The fact is that the participants in the experiment created a full-fledged pedigree for the fictional character: it begins with ceremonial portraits of the great-grandfather and great-grandmother - the Count and Countess Belami. Judging by their hairstyles and robes, they are similar to the contemporaries of the Sun King - Louis XIV. This is followed by their children - marquises, barons, a lady in blue, who, judging by the family tree,for some reason is the wife of Cardinal Belamy and mother of Edmond … In total, there are 11 portraits. If you auction off this entire collection, then, given Edmond's success, you get a decent amount. But the programmers have not yet shared their commercial plans.
COMPETENTLY
Neural networks make diagnoses more accurate than doctors
We asked the head of the laboratory of neural systems and deep learning at the Moscow Institute of Physics and Technology (MIPT) Mikhail Burtsev to comment on the results of the intellectual and artistic trades :
Mikhail Sergeevich, help assess the scale of the event: is a work of art created by artificial intelligence a breakthrough comparable to the victory of the Alpha-Go program over the world champion? Or something else?
- In fact, this is just a good PR move. First, for my taste this painting is a dubious artistic achievement. Secondly, the algorithm was created a long time ago and it can be used for more interesting things.
For example?
“These algorithms are well proven in terms of generating photorealistic images. For example, an experiment is known when realistic portraits of non-existent celebrities were created using GAN. You saw a paradox: an unfamiliar face with recognizable features. The same algorithm can be used to improve the sharpness of the image when there is a request for super resolution. Thanks to artificial intelligence, you can twist the small details of photos to incredible clarity. For example, the fall foliage under the feet of a girl in the park can be generated to such a resolution that every vein of every leaf is visible. It doesn't matter that this does not exist in the real photo, this fragment will be fictional by the network and indistinguishable from reality. In addition, it is possible to improve the quality of blurry and spoiled photos in this way.
The algorithm created these realistic portraits of nonexistent celebrities.
How does the image itself appear on the media? The neural network has neither arms nor legs
“I don’t know how they did it in this particular case with Emond Bellamy - maybe they just printed it out. But I remember there was a project where they painted a painting by Rubens, where the image was made using a robot that painted the canvas with real paints.
- Can you explain to the general reader what a generative adversarial network is
- This is a machine learning algorithm, where one neural network teaches another. One network has the task of synthesizing an image. We load data into it and make pictures based on them. And to train it, we use another neural network that can distinguish between pictures created by a machine and real ones - for example, drawn by a person. And then the adversarial principle turns on: the generating network is forced to look for options to create images that will not differ from the works of artists.
Besides playing Go, are there any areas of activity where artificial intelligence has outstripped humans?
- There are good results in image recognition for all kinds of medical applications. For example, artificial intelligence recognizes X-ray images better than a person, therefore, it makes a diagnosis more accurately. There are also successes in speech recognition. Now the neural network recognizes human speech as well as people. This is a step towards creating virtual assistants, voice assistants. We in the laboratory are engaged in the creation of smart dialogue systems, but it is still too early to say that they can maintain a conversation at the level of a living interlocutor. Although for routine tasks - answers to frequently asked questions over the phone, automatic algorithms provide quality comparable to a person. But we need to weigh which is better: hang on the line for 10 minutes, waiting for the operator, or get a quick answer to a typical question from the machine.
But how can a person who called a call center determine whether artificial intelligence can solve his problem or whether a live operator is needed?
- The principle of operation is as follows: the voice of the caller is translated into text, then the text is fed into a special classifier, which recognizes whether the algorithm can answer this or not. If we have an answer to this question, it will be issued. If the system is not sure what will answer the question, then it is transferred to the operator.
YAROSLAV KOROBATOV