Generative Models and the Concept of Creativity
The question of what is creativity is an enormously fascinating and at the same time very complex one. Although, it got progressively more into focus when it comes to generative models. Machines "creating" music, images, poems and much more let us questions the concept of creativity, which was defined as a genuinely human ability.
Deconstructing the learning, while (more or less) imitating the brain structure should make us rethink what learning in its fundamental neural basis is and how it differs from its machine imitation. Can we actually see it as an imitation? Or should we regard it as an independent technology, having nothing in common with the creativity in humans?
Creation and creativity
When algorithms start to "learn" to create art, they decompose it and reveal the mechanism behind it. The process of creation will be fragmented in its fundamental bits. The way how generative models work nowadays ties to the definition of steps of creativity by Jonathan Gottschall: immersion, assimilation and recombination [1]. However, none of these are limited to humans-only capabilities.
Taking another approach, we might split the term creativity into two fundamental types: combinatorial creativity and emergent creativity. The first is addressing the recombination of fixed primitive elements to create new structures. The latter describing the process where new structures or symbols emerge ex nihilo [2].
Generative models tend to work with the first type, recombine existing elements until something emerges, where we are no longer capable of detecting its distinctive parts. A number of works have tried to define generative art [3][4][5]. Generative procedures have a long history in arts that predates the computer by thousands of years [6]. One proposal on defining AI arts by Lev Manovich takes up the approach of the Turing test:
"In fact, we can extend the famous Turing test to AI arts - if art historians mistake objects a computer creates after training for the original artifacts from some period, and if these objects are not simply slightly modified copies of existing artifacts, such computer passed “Turing AI arts” test."
However, Manovich also points out that even if this approach is logical, it is not sufficient, since there is no commonly accepted definition of “art”. Missing this fixed definition, leaves us without a well-defined target and thus we cannot converge. If we cannot distinguish between a human creation of an art piece and a machine one, did this then become art? Is it the output that is defined as arts, or the process? If it is the process, is the act of programming the ML-algorithm art? If so, can we then still define a human-made painting as art? Is it the humanity in the process, which we based our definition on?
When creativity means to come up with something which we would describe as new or innovative,
and generative models like Generative Adversarial Networks using noise to get a certain degree of randomness for creating "new" things -
can we describe the noise as creativity?
Taking up a neurological perspective, the act of thinking is a deterministic, but idiosyncratic activation of neurons. We could therefore derive the hypothesis that the act of getting ideas is a directed search without a target. Not every path we travel (in that sense neural connection we follow) will lead to a solution, or be determined as "good" by the prefrontal cortex. We might also think of it as a directed search, where one even has indeed a target. However, this target remains changeable and flexible. Machines, however, follow the rule they are programmed with and are not self-determined. An algorithm will, even if probabilistic, react to a given input and converge.
Experiment
Not really to answer, but rather to play with these thoughts, I made a little experiment: First, I trained a Deep Convolutional Generative Adversarial Network (DCGAN)[8] with the images of abstract art1 After generating several new images of abstract art, specific ones were selected. This selection was executed by a biological neural network, with a pre-training of 30 years. The third and last step was to translate the digital images to real-word paintings by using another neural-network-based generator which is connected to motor output.2
Examples of this algorithm are shown in the Figure below.
In essence, the JENNI algorithm (Joint Enhancement using Neural Network Imagination)3 is a process using the selected output of a DCGAN trained on abstract art to generate real world painting.
Reflection
When the painter basically copied the output of a generative model, creating artworks, is this art? To answer this, we might fragment this process of computer-enhanced abstract art creation and think about its separate parts involved in this process. Was the painter the creative entity? If so, because of painting, or because of having the initial idea? Was it the machine involved in that process, because without it, it would not have resolved to the same result?
Addressing the result, Ed Newton-Rex stated: "If the artwork can touch us emotionally, does it matter who created it? Further, for him creativity is much more than just the output- It is the people behind, which we admire[9]. With this in mind, returning to the above-described experiment: Is this art? Is this creativity? Can the painter be admired?
The introduction of the machine in the process of creating something "new", raises further questions about the process as such. Decomposing the procedure of computer-generative art, let the focus shift towards the process of creating an artwork. Following McCormack [10], creating an artwork requires having a certain degree of independence and autonomy. This lets us make a distinction between two terms of generative art: The first one, which is termed "strong", gives creative autonomy and independence primarily to the computer, minimising the creative signature of the human who designed the system. The "weak" generative art, in turn, uses the computer as a passive tool or assistant, where the artist primarily has the responsibility.
Following this distinction, we can regard art, created in a machine-enhanced way, term as weak generative art. However, looking at the intermediate state of the algorithm-produced output, while leaving out the follow-up steps, might be in turn considered as strong generative art.
Conclusion
In essence, generative models let us rethink the process of creativity. And while working on new models, solving upcoming problems in our society, we should also remember to once in a while stop and think about the influence of AI and the thereof emerging responsibility.
Footnotes
1https://www.kaggle.com/greg115/abstract-art
2In other words: Human-painted images with an up scaling in pixel and size.
3Code available online under: https://github.com/jayflyaway/JENNI
References
[1] J. Gottschall, “Edge the rise of storytelling machines.” [Online]. Available: https://www.edge.org/response-detail/26068
[2] P. Cariani, “Creating new informational primitives in minds and machines,” in Computers and Creativity, J. McCormack and M. d’Inverno, Eds. Springer Berlin Heidelberg, 2012, pp. 383– 417.
[3] M. A. Boden and E. A. Edmonds, “What is generative art?” Digital Creativity, vol. 20, no. 1-2, pp. 21–46, 2009.
[4] P. Galanter, “What is generative art? complexity theory as a context for art theory,” in In GA2003 – 6th Generative Art Conference,2003.
[5] J. McCormack and A. Dorin, “Art, emergence and the computational sublime,,” Second Iteration: conference on Generative Systems in the Electronic Arts (Alan Dorin, ed.), CEMA, Melbourne, Australia, 2001.
[6] A. Dorin, J. McCabe, J. McCormack, G. Monro, and M. Whitelaw, “A framework for understanding generative art,” Digital Creativity, vol. 23, no. 3-4, pp. 239–259, 2012.
[7] L. Manovich, “Defining ai arts: Three proposals.” [Online]. Available: http://manovich.net/index.php/projects/defining-ai-arts-three-proposals
[8] A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,” 2016.
[9] E. Newton-Rex, “Can computers be creative?” [Online]. Available: https://www.london.edu/think/can-computers-be-creative
[10] J. McCormack, O. Bown, A. Dorin, J. McCabe, G. Monro, and M. Whitelaw, “Ten Questions Concerning Generative Computer Art,” Leonardo, vol. 47, no. 2, pp. 135–141, 04 2014.
