An artistic creation, made with artificial intelligence systems in Chile, will be presented this Thursday within the framework of the III International Seminar on Audience Development, organized by the Ministry of Cultures.
The people in charge are Denis Parra, a computer researcher and academic at the Catholic University, and Rodrigo Cádiz, an academic from the Institutes of Music and Electrical Engineering of the same university, which form the CreativAI Lab UC, entity that will offer the Intelligence workshop Artificial for creativity at 3:00 p.m.
In this case, elaborated with the Stylegan2 system, a generative model network (or in English GANs, Generative Adversarial Networks) for the creation of images, the CreativAI Lab UC piece is based on the "style transfer" of the online catalog of works from the National Museum of Fine Arts, where the system was trained and learned from this universe of images. Furthermore, the music was also generated by this AI.
Image generated from a series of piano rolls, where an initial image A is indicated, a final one F, and the neural network generates 4 piano rolls images to interpolate between A and F. In this way, more extended audios can be generated than the default output of 128x128. The image is part of Manuel Cartagena's thesis, "Exploring symbolic music generation techniques using Conditional Generative Adversarial Networks".
Origin of the work
Can an Artificial Intelligence system create a work of art? This is the question that has been around for years in artistic, academic and scientific circles due to the progressive progress in the development of Artificial Intelligence (AI) systems and their use as a creation tool in various artistic disciplines. In recent years, novels, poems, songs, and paintings have been developed from extensive AI research around the world.
Parra has been researching personalization systems for ten years, which use artificial intelligence to recommend musical and visual art, while Cádiz has a recognized track record in musical composition, electronic music and instrument creation.
About three years ago, they decided to start a collaboration based on an interdisciplinary project of the UC Vice-Rector for Research (ArTeCIH program) with the purpose of exploring the possibilities of artificial intelligence as a tool for artistic creation. This is how the CreativAI lab UC was born.
Cádiz points out that the particular work presented at the International Public Development Seminar is the result of Manuel Cartagena's master's thesis, who is studying ways to use GAN-type networks (Generative Adversarial Networks) trained with images to generate music in symbolic format (as if it were a score).
A GAN-type network is capable of generating signals (images, music, text, etc) simply from noise based on training, where the neural network is “shown” a large number of signals that it wants to receive. the network learns. It is one of the types of generative networks, he says.
Images of the face-generating GAN network, trained with paintings from the Chilean National Museum of Fine Arts.
How to Add or Remove 'Turn off BitLocker' Context Menu from Drives in #Windows7, #Windows8, and #Windows10 https://t.co/WdxrTGocMH
— TenForums Fri Jan 22 18:33:21 +0000 2016
Methodology
Parra comments that to make the music video that they are presenting at the seminar, they trained two generative neural networks: one that generates faces and another that generates music in symbolic format.
"For the face generating network, we collected a dataset of paintings from the MNBA and processed it to be left with only faces. We used this dataset to train a GAN-type neural network that generates 'new' faces inspired by existing faces in paintings, engravings, sketches, and photographs of the MNBA," he explains.
Cádiz adds that "to generate the music we used another GAN network. Starting with a neural network architecture of the StyleGAN2 type, we proceeded to train it with images of piano rolls, a basic form of musical representation. Then the network learned to create images in piano roll format, which when translated into music, result in new, artificially created musical material".
Finally, to create the video, they used audio generated by the music GAN network as input to the neural network of faces. This made it possible to go "taking a walk" so that the visual GAN network generates faces when listening to the input music, adds Parra.
World Background
Parra states that, from the point of view of computer visual generation, different algorithms have been used since the 1970s to create digital art.
However, the so-called generative adversarial networks, created in 2014 by the computer scientist Ian Goodfellow, have had a dizzying development that has led them to be used to generate faces of non-existent people that seem real, and different works such as " The portrait" of Edmond Belamy, which sold for more than $400,000 at Christie's auction house in late 2018.
In terms of generating music with these methods, there are many precedents, the most famous perhaps being Magenta, a Google initiative, where research is carried out aimed at generating music with artificial intelligence. "In terms of applying this specific network, StyleGAN2 to the music generation, we don't know of any other precedents," says Cádiz.
What is the future of these initiatives?
"I am of the opinion that just as in the past it was feared that photography would kill off painting and ultimately ended up promoting new forms of visual art such as abstract painting and photographic art, artificial intelligence will help increase human capabilities for creativity and create new forms of artistic expression," answers Parra.
"In the future, I have no doubt that we will attend concerts where the music was composed entirely by machines", complements Cádiz.
"We are in the transition to that stage, where we are shaping what the future of computer-aided music creation will look like. From a perhaps more philosophical point of view, it is fascinating to see how our understanding of why these networks are able to doing what they are will illuminate our understanding of the human mind and how it is capable of finding creative solutions to the problems we face as a species."