The art project of Orkhan Mammadov, an installation with realtime AI, generating traditional patterns.
I have covered technical development, programming, GAN training and overall setup of the system. I've used Progressive GAN as AI and vvvv as main integration tool. The installation has run at the Venice Biennale for 4 months.
<b>Official info ::</b>
The installation’s visual component imitates traditional Azerbaijani patterns. As a result, viewers see how AI builds new alternatives. These alternatives have a synthetical nature that has nothing to do with the history of authentic ornament patterns. AI imitates and fakes a traditional learning process usually handed down from generation to generation. By accessing the data of authentic ornament images, AI becomes an independent master that can invent new ideas to update culture.
The installation’s visual component imitates traditional Azerbaijani patterns. As a result, viewers see how AI builds new alternatives. These alternatives have a synthetical nature that has nothing to do with the history of authentic ornament patterns. AI imitates and fakes a traditional learning process usually handed down from generation to generation. By accessing the data of authentic ornament images, AI becomes an independent master that can invent new ideas to update culture.
<b>Technical description :: </b>
Recent years have marked totally new approach in the utility computing, generative methods, digital production, and somehow in new media in general. Here came Artificial Neural Nets - computing systems, developed and constructed on the principles of natural biological neural systems. Instead of processing some complex step-by-step algorithms, these systems process vast amounts of data with millions of simple operations, optimizing their internal parameters (neurons) in accordance to those data. Such process is called 'training' or 'learning', and it quite closely resembles the human studies with the gradual appropriation of the knowledge (hence another name of this technique - Deep Learning).
The fundamental power of Neural Nets as a computing tool is in its ability to find correlations and patterns in the very different types of data. So it's quite reasonable that such method can also generate real visible patterns - from the basic modern geometry to the complex medieval filigree.
The most popular, promising and evolving Deep Learning architecture nowadays is called GAN (Generative-Adversarial Network). It consists of two cross-linked neural networks, one of which tries hard to create fake imagery and another one tries to guess if it's fake or real. Both networks share their guesses with the opponent, so that both could learn from the process, operating better and better.
For our goal we took one of the best available GAN architectures - ProGAN (Progressively Growing GAN) by NVidia. It learns the imagery from the dataset, progressively going from the lowest downscaled resolution (4x4) to the full scale pictures, ensuring the proper attention to details on all levels. We have fed it with the dataset - decent amount (tens of thousands) of the authentic patterns (again, from simple ones to quite excessive combinations). After about a week of training process, the system was able to produce new patterns - reminiscent of the original imagery, yet different and ever-changing.
The system's output - realtime generated visuals - was mapped to the solid LED ring with a complementary information panel, illustrating the whole process happening inside the server. Imagery creation is also integrated with the sound (generated realtime as well) to emphasize living nature of the visuals.
Recent years have marked totally new approach in the utility computing, generative methods, digital production, and somehow in new media in general. Here came Artificial Neural Nets - computing systems, developed and constructed on the principles of natural biological neural systems. Instead of processing some complex step-by-step algorithms, these systems process vast amounts of data with millions of simple operations, optimizing their internal parameters (neurons) in accordance to those data. Such process is called 'training' or 'learning', and it quite closely resembles the human studies with the gradual appropriation of the knowledge (hence another name of this technique - Deep Learning).
The fundamental power of Neural Nets as a computing tool is in its ability to find correlations and patterns in the very different types of data. So it's quite reasonable that such method can also generate real visible patterns - from the basic modern geometry to the complex medieval filigree.
The most popular, promising and evolving Deep Learning architecture nowadays is called GAN (Generative-Adversarial Network). It consists of two cross-linked neural networks, one of which tries hard to create fake imagery and another one tries to guess if it's fake or real. Both networks share their guesses with the opponent, so that both could learn from the process, operating better and better.
For our goal we took one of the best available GAN architectures - ProGAN (Progressively Growing GAN) by NVidia. It learns the imagery from the dataset, progressively going from the lowest downscaled resolution (4x4) to the full scale pictures, ensuring the proper attention to details on all levels. We have fed it with the dataset - decent amount (tens of thousands) of the authentic patterns (again, from simple ones to quite excessive combinations). After about a week of training process, the system was able to produce new patterns - reminiscent of the original imagery, yet different and ever-changing.
The system's output - realtime generated visuals - was mapped to the solid LED ring with a complementary information panel, illustrating the whole process happening inside the server. Imagery creation is also integrated with the sound (generated realtime as well) to emphasize living nature of the visuals.
Concept & Design: ORKHAN
Software Development: Vadim Epstein
Generative sound programming: Duganov Ilya
Curators: Emin Mammadov & Gianni Mercurio
Organization: Heydar Aliyev Foundation
Premiered at: La Biennale di Venezia, 2019
Software Development: Vadim Epstein
Generative sound programming: Duganov Ilya
Curators: Emin Mammadov & Gianni Mercurio
Organization: Heydar Aliyev Foundation
Premiered at: La Biennale di Venezia, 2019
