(ART)ificial Intelligence
By Dan Kershner, Zach Brill, Annan Miao, Troy Schwab
Motivations and our ‘Why?’
We were spurred on by the question of the relationship between art and computer science, and even more so considering how models in artificial intelligence can relate to concepts of creativity and innovation. As Dennett suggests (Haugeland 2000), intentionality is an important question when approaching artificially intelligent systems. Our team, made up of undergraduate computer science majors, certainly wouldn’t be able to produce a system with original intentionality, though if a human perceived an artificial system as having original intention, then for our purposes, that prescribes a level of authenticity to our system’s output.
We stumbled upon a class of AI models called GANs — generative adversarial networks. This paper will introduce an approach to creating images and even art pieces using these models, considering a tensorflow dataset of images (MNIST and CIFAR10) as training data. The manipulation of the custom image set wasn’t fully achieved in the allotted time as a result of mode collapse, though the functionality with CIFAR10 and MNIST sets function properly.
Implementation
Our implementation generally focused on a minimax function in Figure 1. The function essential is a game between the generator network and the discriminator network. The goal being for the generator’s accuracy to be maximized and the discriminator’s error found to be minimized. Eventually, a GAN using this function will optimize itself, where the images can no longer be perfected, and the model is fully trained. It was our attempt to understand the theory and structure of the GAN.
We begin by taking a random noise vector and sending it to the generator network, and letting it alter it in any way through its trainings from the original data set, and then sending it to the discriminator. The discriminator will then fail the image if it is able to detect a difference between the generator’s output and the original training set. Both of the networks are convolutional networks, so essentially their analysis concerns the color gradients of the images, and as such can approximate the form, shading, and even stylistic choices (if the input is a work of art) that are resultant of gradients. Keras is our high level network to help develop and evaluate deep learning models with tensorflow as our backend. We use matplotlib for plotting and then tqdm to show a progress bar for every epoch. We further considered a Leaky ReLU algorithm as our activation function, but ended with a sigmoid function to lead our output to a singular number.
Ethics
The incorrect usage of this powerful network can create very extreme ethical dilemmas. Most noticeably, the open source project of Robbie Barrat, an artist who works with artificial intelligence. He developed a GAN model to produce abstract portraits around 2015, which was used by the French art collective called Obvious. This painting was then sold for nearly half a million dollars, excluding Barrat, and Obvious claiming they used his code as inspiration and modified it slightly. In Figure 2. we can see the painting which Obvious made, and in Figure 3, we see Barrat’s paintings from his algorithm.
Extending from this controversy, we need to consider if the art belongs to anyone in particular? The creator of the algorithm? The creator of the specific code? Perhaps the art belongs to all the artists that were enveloped in the training data. Following this consideration, does the fact that the code is open source affect the ownership? Beyond that, is it even possible to ethically sell one of these works with these questions in mind?
With these many questions, we will dilute the possibilities by setting a foundation of yes, the algorithm and code was stolen. The main question we attempt to now answer focuses on the rights of open source algorithm development, and the profiting of from use of that open source algorithm. In kantian ethics, one may be said to have acted morally if the rules used to do so may be considered a moral law. In this situation, the theft and profit of another’s work, even though it was offered freely, is wrong, as it was only offered freely on platforms like github to aid other artists in their experimentation and research. Therefore, from the perspective of a kantian approach, it is clear that this is unethical.
Though, let’s assume Barrat was never aware of this utilization of his software. He then is never hurt by the profits that Obvious made, and the software publicly is still maintained and aiding other artists with experimentation and research. His well-being therefore does not decrease at all, while Obvious, who receives fame and economic profits experience an increase of well-being. This act then would increase the overall well-being of all parties involved. Thus, from the standpoint of a utilitarian analysis, it is fair to say that no immoral act occurred and this was perfectly ethical.
The main distinction in both of these analyses is that the utilitarian analysis equates happiness (usually offered as the metric in utilitarian analysis) to well-being and also assumes ignorance from Barrat. Therefore, it is fair to say that the kantian approach is more effective in analyzing this dilemma and that Obvious’s utilization of Barrat’s code was unethical.
Philosophy
With this model now creating gorgeous art works, we must consider philosophical consequences of AI generated art. Who is the artist in these cases? Can this be thought of as a creative endeavor? Does this generation of works function within the art world?
Barrat relates his role often to the artist Sol Lewitt (Bailey 2018), who would write out instructions or rule sets for creating art that others would execute. Barrat simply programs the machine, offering it boundaries in how it operates and then provides the training set to see what it does. It can be thought of synonymous to a person who has only ever seen Monet paintings grabbing acrylics and copying the gradients of colors. In this case, the person simply learns from Monet, and is bound by the canvas as well as acrylics and knowledge of the color gradients. Thus, the role of Barrat can be thought of almost as a conductor, where the performing role belongs to the GAN. The GAN is not independent as it is simply a tool, identical in utility to markov chains in poetry, or random granular synthesis in music. It is up to the conductor in how to manipulate (or not) the output to some other form.
Framing the GAN as a tool removes the questioning of its creativity, but for the sake of discussion, we will consider the GAN as the artist for a brief moment. In this context, the GAN may be said to have creativity, but only in a short term perspective. This being a result of only a basic “grasp” of the color gradients. There may be no innovation over time that a human artist experiences that requires true experimentation and progression. Only information provided by the training set is knowledge that the GAN has and can have. It can never grow beyond the bounds of that set. This clearly demonstrates original versus derivative intentionality, as the intent of the colors chosen pixel by pixel is derived from the knowledge gained from the training set, though the model may never offer an original intent in a painting. Similar to Searle’s Chinese Room Argument (Haugeland 2000), our GAN is restrained to the bounds of the room (being the training set) and the forms of the letters (being the pixels), but at no point in time, may the GAN alter those restrictions, or even understand the meaning of them.
Returning to our prior conclusion that a GAN is simply a tool, how does this fit in the art world and coincide with art (and aesthetic) theory? W. E. Kennick has held less strict rules on that which is or is not art (Kennick 1964). He likens art to being something which is considered art, as such art is not art without the audience or the creator who would need to have the perception of it. Though with the interaction provided by the developer creating the GAN and making the images, would be enough to qualify an audience member, and as such the work as a normal piece of art despite the means by which it was made.
Conclusion
What worked
Our project was a general success; our implementation worked properly and we found interesting results from the input image sets while learning how to properly utilize libraries to their potential to minimize our work. We also encountered some fun and interesting generated images when we trained the discriminator only on images of birds, for example. When we used this approach, our results were quite unique around ~10 epochs into training. This is because the generator was attempting to produce images of birds, but in practice it was generating images that mostly had some indication of a blue sky and a winged object. To the untrained eye, if these were enlarged and clarified they could have been mistaken as abstract paintings of a sort. This was about as close as we came to generating something that we would consider rudimentary art.
What Didn’t
The difficulty came when attempting to implement our own training image sets, due to the fact that unexpected mode collapse had occurred. We deduce that this is a fault in our image manipulation prior to analysis by the GAN, as our model functions properly on the built in image sets. We also encountered extreme PC slowdown once we began manipulating our own data set (a large image set of photos of flowers.) We were able to resize all of the images to be the same dimensions after which we converted each image to a numpy array as this enables easy training in Keras. However it was after we began to run our GAN on this data set that we encountered what is likely hardware problems. One member even had to reinstall their operating system due to some extreme overworking of their hard drive that persisted even after multiple restarts.
Takeaway
It was particularly interesting and ultimately the most rewarding to analyze the ethics and philosophy of our project, especially noting the events between Robbie Barrat and Obvious. The philosophy questions more broadly considered the ripple effect that computer science and artificial intelligence may cause in external fields. Though, with heavy application of our in-class readings derived from Haugeland’s Mind Design II as well as articles from different sources online, we find ourselves excited to see computer science become implemented more often and to grander extent within external (particularly more creative humanities driven) fields.
Resources
Haugeland, J. (2000). Mind design II: Philosophy, Psychology, Artificial Intelligence. Cambridge (Massachusetts): MIT.
Cohn, G. (2018, October 25). AI Art at Christie’s Sells for $432,500. Retrieved from https://www.nytimes.com/2018/10/25/arts/design/ai-art-sold-christies.html
Vincent, J. (2018, October 23). How three French students used borrowed code to put the first AI portrait in Christie’s. Retrieved from https://www.theverge.com/2018/10/23/18013190/ai-art-portrait-auction-christies-belamy-obvious-robbie-barrat-gans
Karras, T, et al. (2018). Progressive Growing of GANs for Improved Quality, Stability, and Variation. Santa Clara (California): NVIDIA.
Huang, H, et al. (2018). An Introduction to Image Synthesis with Generative Adversarial Nets.
Bailey, J. (2018, April 5). AI Art Just Got Awesome. Retrieved from https://www.artnome.com/news/2018/3/29/ai-art-just-got-awesome
Zuo, Y, et al. (2018). Generative Adversarial Forests for Better Conditioned Adversarial Learning.
Goodfellow, I, et al. (2014). Generative Adversarial Nets
Amos, B. (2016). Image Completion with Deep Learning in TensorFlow. Retrieved from http://bamos.github.io/2016/08/09/deep-completion/#ml-heavy-generative-adversarial-net-gan-building-blocks
Kennick, W.E. (1964). Theories of Art and the Artworld: Comments. Journal of Philosophy, Inc. Retrieved from https://www.jstor.org/stable/2022938
