One can hardly fail to notice that, from Benjamin’s chef d’oeuvre in 1935, The Work of Art in the Age of Mechanical Reproduction, derives the book title Art in the Age of Machine Learning. While the aura of artwork is dimmed with the invention of photography, the birth of machine learning allows us to reproduce not only an artwork’s content, but also the procedure by which an artwork is produced in the first place. Given that the development of technology is undeniably entering a new phase, the reality today may elude the critiques that have been made in the last century. Owing to the complexity of computer technologies such as algorithm, regular internet users could by no means understand how they function. Their ignorance consequently leads to exceeding admiration and fear. On the one hand, they are amazed by the omnipotence of these unknown beings, thus crowing searching engines as “the Great Gods.” On the other hand, they worry that Midjourney and DALL-E will eventually bring about the unemployment of artists and designers. How, then, should we interpret the relation between art and technology in our time? A short introduction to the field and a few case analysis will probably shed lights on this topic.
Art in the Age of Machine Learning is a book for beginners. The basic concepts in machine learning is clarified through an abundance of explanations of machine-generated artworks. The author, Sofian Audry, an artist and a scholar, is currently a professor of Interactive Media within the School of Media at the University of Quebec in Montreal (UQAM). It is not difficult to notice from his academic career how he is capable of making such a complex subject accessible to non-professional readers: majoring in computer science in his undergraduate study, focusing on machine learning and interactive media in graduate school, and eventually acquiring his doctoral degree in Humanities from Concordia University. Yoshua Bengio, a precursor in deep learning and the winner of Turing Award in 2018, was Audry’s Master advisor, who also contribute to the foreword for his former student’s work.
Audry started out with some misunderstanding of art after a decade of study in information engineering. But gradually, he finds out that art is more of an activity concerning “raising questions” than “solving problems,” and the latter characterizes the central consideration in the field of engineering. The entire book is divided into three parts: training, model and data. These three parts also coincide with the three elements in machine learning. To quickly summarize, machine learning requires a certain amount of data in order to be trained, and, from this process generates a functional model. However, before he starts to explain how machine learning works in actual artworks, Audry points out four common myths prevalent in the realm of machine-generated art:
Myth One: Artificial intelligence, machine learning and deep learning are the same thing.
They should be understood as three things located in three different scales. Artificial intelligence covers multiple machine-learning approaches including algorithm and self-training machine learning. Among them, there is also deep-learning, an approach whose training system mainly relies on the technique of artificial neural network.
Myth Two: Machine learning is a new thing.
The concept of machine learning can be traced back to cybernetics developed in the 1940s. Although the term itself, along with another famous term artificial intelligence, did not appear until the 1950s. As for machine learning in art production, it is more difficult to tell from which period it was born, since most works adopt this term in a rather metaphorical way than actually applying the technique to their creation process.
Myth Three: Artists are totally absent in machine-generated artworks.
Despite some astonishing outcomes that the system of machine-learning may bring about occasionally, human “labors”, such as building large-scale databases and modifying training algorithm, are still necessary to most cases in this book. More importantly, even when computers bear a few responsibilities to make choices, there are decisions only artists can make.
Myth Four: Machine learning will give rise to superintelligence very soon.
From the author’s point of view, this opinion is more a delusion of technology in its entirety than a myth, just like the futurists in the early 20th century, who believed that humanity will be replaced by mechanical technology.
In addition to breaking the myths, Audry illustrates the fundamental contradiction people would encounter when applying machine-learning technique to art production, and some possible solutions. For instance, in the realm of machine learning and computer science, optimization is a common and classic method. However, this method cannot be applied to art production, because art, as an activity indifferent to solving problems or offering answers, whose meaning is acquired only in a background or in a context, can never be optimized. It is precisely because of the opposition between the two, that is, between art and a highly automatized system, the choice between different models becomes a concern for the artists who must pick suitable visual or phonetic effects for the work. Their intervention in the system is like an act to “bend” the one-way training process so that the need of aesthetic element essential to the artworks can be met. It could be giving immediate feedbacks to disturb the training process or building an objective function to affect the system through mediation so as to observe its reaction. Both methods treat the training process as key to aesthetic experience.
However, the idea that one should comprehend thoroughly what happens in the whole training process is not always realistic, for traditional programming is different from machine learning. Generally speaking, programs are designed and written completely by engineers. Data are input into the program and then the program outputs the outcome. In this sense, the engineer is able to observe simultaneously the input and the output. In contrast, machine learning enables collected information to produce a program after the information undergoes a learning process. That is to say, the program is not completely produced by human being. Such a trainable program is called “model”—a mathematical representation of the real world.
With an attempt to explain how such a machine training process can be accepted as a way of art production, this book also examines the interrelation between the development of contemporary art and that of artificial intelligence. On the one hand, from the 1960s to the 1970s, contemporary art was defined more by how a work is produced than by what a work eventually is, in which case the producing process of the work and the conduct of the artist are accentuated. On the other hand, based on the artificial intelligence which develops from cybernetics since the 1940s, bottom-up AI like nouvelle AI appeared in the 1980s. Likewise, roboticist Rodney Brooks in his essay “Elephant don’t play chess” published in 1990 argues that the primary source of machine learning should come from external environment. A robot does not imitate chess players who move their pieces according to the rules at hand. A robot should behave like a wild animal who, struggling for its survival, must be sensitive and responsive to the nature.
The birth of these self-organizing robots, strange as they seem, has implicitly reversed the role of human beings and that of the machine in art production. The Three Sirens, a band of robot, is a case frequently cited in this book as an exemplary. Artist Nicolas Baginsky reported that, after setting basic parameters, he let the robot keep “stimulated” by the data received from the real world. In this case, the machine is no longer human being’s intelligent assistant; conversely, it is the human being who becomes the robot’s production assistant.
The last part of this book focuses on the material of machine learning: data. A great amount of data is indispensable for the stage of model training. Thanks to the rise of the Internet, megacorporation nowadays can easily collect massive information from Internet users, which was impossible a decade ago and made these large companies the most advantageous organizations to develop the machine learning technique. However, artists cannot collect data in the same way, which is another difficulty when applying machine learning to making artworks apart from the necessary skill of programming. Audry gives his readers a few examples to elucidate possible methods formulated by artists: First, Laetitia Sonami and Suzanne Kite, with their flight records, use the training model of Wekinator, an open-source software, to create their own music instruments. In a similar way, Sougwen Chung, with her paintings drawn during the past two decades as data, trains the robotic arm DOUG to collaborate with her in new creations. As for Anna Ridler, she accumulates 10,000 photos of tulips she took during blooming seasons to train generative adversarial networks. Finally, artists like Brian House produce their works by “outsourcing” the process of data collection. In his Everything That Happens Will Happen Today (2017), he invites anonymous volunteers to record in an application their itineraries in New York City.
To sum up, even if mathematical formulas are practically absent in Art in the Age of Machine Learning, for a layperson in the field of machine learning like me, it still takes effort to read this book. Besides, the criticism offered in this book seems incompetent to change the existing condition. In Chapter 8, the author shows his suspicion against plans like artist-in-residence or sponsorships supported by tycoons of technology companies, because exploitations are not uncommon when these companies try to get some ideas from the artists for their machine learning researches. The means of production and the knowledge of creation seem to be alienated gradually from creators and producers, which is a severe problem beyond any investment in equipment update.
Sofian Audry | Art in the Age of Machine Learning：https://vimeo.com/677506809
LEE Chia-Lin graduated from the Department of Foreign Languages and Literatures, National Taiwan University, and the Institute of Contemporary Art & Social Thoughts, China Academy of Art. She is now pursuing a Ph.D. in Fine Arts at Taipei National University of the Arts. Her research focuses on the culture, media and art developed and created in the digital era. As the founder of ZIMU CULTURE, LEE also works on curatorial projects and publishes books.