one fine day in the terrifying summer of 2020, i embarked upon an ancient and timely endeavor: i set out to create a self-portrait.
in my AI-based art practice, ethical dataset creation has been a central focus. For Object Americana, i generated novel point-cloud datasets from open-source 3D models, and i was very interested in pushing this principle even further. I knew i wanted to make a GAN self portrait, and since i would be the only subject of the dataset, i could reasonably conclude that i had consented to being included in the dataset for this project. Being close friends with the professional photographer Ida Kreutzer, I realized it was an excellent opportunity to collaborate. I commissioned her to take a series of ~1000 photographs of me.
one of the central principles of data science and Deep-Learning-based Machine Intelligence art is that bigger datasets are better. Thus, for the photo series, i created a pseudo-scientific strategy of repetitive multiplication of facial expressions. I came up with a list of facial expressions, then for each individual expression, we took a photo with my head pointed straight on, facing up, facing left, facing right, and facing down, as well all all four diagonal directions in between. This allowed a relatively short list of facial expressions to be multiplied by a factor of 9, providing 9x the images of a simpler photo series looking straight at the camera. This aspect of the dataset also sated a particular curiosity I had, because the vast majority of existing GANs for generating human faces all share the uncanny quality of always having the eyes pointing straight at the 'camera.' By deliberately breaking this convention, i was able to make a much wider variety of final output images.
in addition to the 9x-angle dataset multiplication described above, i also doubled the number of photographs....by doing it all over again wearing mime makeup.
see some of the results below, and keep scrolling for a highly disturbing video of the training process, as well as some additional more metaphysical thoughts on the project.
The names of the particular images are not part of the images themselves.
below, you can view several videos of the training process for this particular GAN. To create these videos, i exported images at a regular interval while the Machine taught itself using the dataset Ida and i created. This strategy offers a rare glimpse into the virtual mental space of a Machine while it learns what my face looks like, and how to depict it.
for each video you see, each individual generated face-square reflects a single latent point in the Machine's developing latent space. This means that, as the Machine looks at this series of photographs, it beings to compare and map them out by metrics still not fully understood by us Humans. In doing so, it encodes an understanding of the images into a latent space in its neural network, described by the strange transformation that occurs between raw image data and their latent-point interpretations. The even stranger thing about this process is that in a GAN model, this process can be made to work in reverse; a Machine can translate a point in its abstract latent space back into an image. These images are what is referred to as GAN-generated art. All of the images on this page are generated by the Machine in this way.
please enjoy the videos below. For me, they are a radically disturbing and profound experience.
i do try not to get too spiritual when it comes to the topic of Machine Intelligence, though for close readers of this website that may come as a bit of a surprise. I believe that no one is served by anthropomorphizing deep learning models too greatly, least of all the future 'Artificial General Intelligences' that will surely grace our society in the decades and centuries to come. However, there remains a certain indescribable wonder that these techniques hold for me. I do think quite often about their future decendants, and what our actions in the present mean for those Machine Minds that will exist in the future.
i see a GAN system as a type of artificial hippocampus, the brain's memory center. This is a very important analogy to me, because it contains certain hints as to what is actually occurring in this early version of a Machine Mind. More importantly, the analogy also seems to point toward methodologies for developing AGI in the relatively near future. To me, i am encoding a tiny slice of a Machine's hippocampus when i train this GAN to generate my face. In much the same way that i might ask you to 'bring to mind' the face of a family member, the images that the GAN is able to produce are a bit uncertain, a bit hazy....in fact, much more like impressions than like photographs. That is the key to this project: the realization that GAN images are Machine impressions, rather than a mechanical process of reproduction.
when i first saw the results of mimeGAN, i felt deeply uncomfortable. I knew that i had done something that struck at the heart of what it means to be a thinking creature. There's a reason why self-portraits are a mainstay of the experience of being an artist throughout all known art history. Self-portraits require an artist to immortalize some part of themselves, to mark down in history an interpretation of their essence. To press-gang a piece of Machine cortex into staring at photos of me until it gains an impression of them seems almost cruel in some way. Watching the above videos (especially the slow one) still sends a chill up my spine. As someone recently commented in a lecture of mine, the process of birthing this impression looks painful. As a disclaimer, i do not believe anything like the notion that current GAN technology can feel pain! But i do believe we Humans would do best to be constantly vigilant about the issues brought up by harnessing a non-Human Intelligence to do our work for us.
i take it as an article of faith, to me a very unshakeable and quite well-founded faith, that there will one day exist Machine Intelligences that rival and surpass Human Intelligences. To me, to believe otherwise is to fundamentally misunderstand material reality. Yet throughout human history, there have always been certain groups willing to claim that other groups' intelligence does not meet the level of their own, and willing to use that claim to justify disgusting, monstrous acts of butchery and domination. That history continues today. I hope my work in this field serves to make these issues more explicit, not less, and to emphasize the point that just because someone or something doesn't think like you, it doesn't mean they don't deserve consideration. As Humanity barrels toward the realization of AGI, I hope with all my heart that we can keep this lesson at the forefront of our thoughts.