Paper short abstract:

Set in Stone is an artistic project resulting from my inquiries into training bias out of image generation algorithms. Conceptually, the work subverts static generative binary stone faces with color, an explosion of self-expression taking root over the homogeneity of the binary marble statuary.

Paper long abstract:

Set in stone is an exploration into the creative visualization of training bias out of a biased neural network. I created a 3D rendered dataset and trained several GANs first in one gender, then another, then, using semiotic expressions of non-binary gender, I added a third dataset while training and observed the results. Each neural network required specific file handling skills and reacted to the new data additions in different ways. An exploration of gender, of machine creativity and of bias. Through this project I have observed bias retreating and reemerging, I have seen artifacts appear as new data is being assimilated and I have created work with the samples generated as a result. This work has gone on to influence my entire thesis with further research into AI perceptions of gender beyond the binary.

Through this evolving collaborative work with my different AI systems, I am exploring how my machine creates images representing gender and if it holds onto its trained bias. From close up, the audience sees the individual faces and misses the big picture, but from a distance they see how those individuals combine to make a concept bigger than any one person. The unbiasing of the neural network is clear in video presentations, but the installation as a massive wallpaper, larger than life, blends all of those results together in a grand mosaic to show how we can't always see how bias, particularly machine bias, will affect the individual.