The-Island-was-a-World_catalogue-digital

8 actually make a computer create like a human artist. After a short reflection, Nees replied that this was possible if only they could provide him with an exact definition of how a human artist creates. This answer upset the professors so much that some departed in a rage. At the beginning of his book, The Creativity Code , 3 Marcus du Sautoy, a mathematician and researcher of creativity and artificial intelligence, adapts the same question to the 2020s: Would it be possible to break the code of human creativity and, more importantly, can it be duplicated to create an artificial artist? This question, central to current techno-cultural discourse, is deeply rooted in human history. It echoes the myth of "the machine as an artist" that Broeckmann describes so well and humanity's relentless aspiration to reach the ultimate Golem algorithm. A project carried out by Elgammal et al. of Rutgers University 4 clearly describes the research approach that aspires to create an "artificial artist." Elgammal’s team developed a unique model named Creative Adversarial Network (CAN), a version of Generative Adversarial Networks, an early AI model. The declared purpose of the project was to create art without the involvement of a human artist yet using human art in the machine learning process. To learn what stands for art, the system trained on 75,753 paintings extracted from Wiki Art, divided into 25 art-historical style categories from the 15th to the 20 th century. While the goal was eliciting images that would be identified as art based on the learned dataset, the model was designed to motivate the image generator to "creatively" deviate from the style categories it had learned. To achieve this, the system's internal feedback component, the "discriminator," preferred works that were difficult to attribute to a specific category, such as Impressionism or Abstract Expressionism. As a result of this feedback, the system learned to avoid imitating the styles it trained on and deviate from them, albeit moderately. In other words, in the Elgammal et al. CAN model, while the definition of art relied on the Wiki Art collection, creativity was defined as deviating from familiar categories. The researchers aimed to mislead the spectators and challenge their distinction between human creations and those of the AI model. Indeed, some respondents identified the CAN works as human. While higher rates of abstract expressionist paintings were identified as human in the experiments, more CAN model paintings were perceived as human than those displayed in 2016 at 3 du Sautoy, Marcus. 2019 . The Creativity Code: Art and Innovation in The Age of AI . Cambridge, Massachusetts: Harvard University Press. 4 Elgammal, Ahmed, Liu, Bingchen, Elhoseiny, Mohamed, and Mazzone, Marian. 2017. CAN: Creative Adversarial Networks, Generating "Art" by learning about styles and deviating from style norms . https://doi.org/10.48550/arXiv.1706.07068