NeuroArch: Architectural Imagery of Artificial Intelligence
Thoughts and practical application of GAN networks [russian version]
Today we are witnessing digital transformation in every aspect of life— otherwise called a paradigm shift to the information society. As I wrote in my program article, this provides us with a completely new toolkit, makes us rethink existing practices, and invent new methods of working with newly emerging technologies and concepts.
However, the technological future puts questions before us, including ethical and humanitarian, about the possibilities and limitations of a human.
Many thinkers are problematizing human interaction with artificial intelligence (AI) and the world of virtuality: “intelligence migrates from protein to sand (albumen VS silicon)” [Efim Ostrovsky].
The explosive growth and penetration of machine learning and artificial intelligence also affects the principles of creativity. If earlier AI was limited to classification and regression tasks, now generative algorithms come to be used to create objects and works of various natures, previously available only to the human mind. Art becomes part of technology, while technology becomes art. In this article I will try to cover more ethical issues than technical ones.
In the era of digital media, visual representation has become the main quality of an object, its digital image. The backside of this phenomenon has become uniformity, since the images have become similar to each other, due to similar techniques and clichés are used for creating them. The abundance of visual information leads to its unification and devaluation.
The idea of the NeuroArch research project was, firstly, to explore the possibilities of generating new images based on existing ones, secondly, to explore the methodology of applying generative machine learning for creative tasks and design, and finally, to give a critical assessment of visual culture:
if algorithms are able to generate new images, similar to the degree of indistinguishability, shouldn’t the AI be entrusted with their creation?
As you know, the essence of GAN networks is that the “generator” algorithm gradually learns to create an image in such a way that the “critic” algorithm cannot distinguish the real image from the generated one. Thus, you can create a new image based on the existing array, as if it were in the same collection. Moreover, it becomes possible to control the parameters of the generated image by changing its properties.
I collected about 6,000+ training samples of examples of modern architecture from popular resources such as Archdaily, some architecture oriented Instagram accounts; among them there were both visualizations and photographs. As a generative algorithm, the architecture of the WP-GAN neural network is used, from the repository for the book by David Foster, “Generative Machine Learning”, which is used to generate faces. You will find a more detailed description of the various GAN algorithms in the book.
Architect and designer Michael Hansmeyer defines generative design as “thinking about designing not the object — but a process to generate objects.”
The designer’s method becomes more like programming an algorithm that defines a framework of the creative process.
Finally, we get not a specific object, but a trained algorithm or a model, capable of generating an any number of new objects.
Analyzing the samples thousands of times (in NeuroArch 110,000 learning epochs), the neural network algorithm learns to extract characteristic features from the samples, forming the so-called a feature space in which features are grouped as model parameters. These features are then used to synthesize new images. The space of features is not discrete, but continuous; it does not consist of separate patterns, due to which a smooth transition between two images is possible.
As a result, the GAN network was able to demonstrate its ability to generate new images. Of course, we are not talking about a detailed picture, but about an impression or architectural fantasy, which did not exist before, but which is woven from many real prototypes. It can be assumed that the generated images have a special aesthetics that makes them related to works of abstract artistic trends such as Cubism or Expressionism.
The question remains, where to draw the line between creativity, as an act of creating something, and technical computation. Just as in quantum physics there is the entanglement principle, when the quantum states of several objects turn out to be interdependent, the role of a person or an algorithm is difficult to isolate from the result. I believe that we are dealing with a new kind of synthesis of man and machine that requires careful comprehension and development of the methodology.
At the moment, we can say we got a powerful toolkit into our hands capable of analyzing and generalizing large data arrays that are inaccessible to humans. Using it, you can penetrate deeper into the essence of things, analyze them and synthesize new ones. I expect that design and art will gain a new impetus for their development, enriching their means and methods.
The author of the publication and graphic materials, unless otherwise indicated: Roman Kuchukov / roma.kuchukov@gmail.com
Facebook: https://www.facebook.com/ICARurbs
