Ersatz Intelligence: Implications of Machine Learning for the Generation and Interpretation of Art and Artifacts

In this presentation we will speak about "Ersatz Intelligence", the title of an artwork and a term that we are using to describe interactions with Machine Learning (ML) and Artificial Intelligence Algorithms (AI) over the course of an artistic project. This project stems from a group initiative entitled Collaborative Artistic Production with Generative Adversarial Networks (GANs) that is currently underway at Hong Kong Baptist University’s Augmented Creativity Lab. The project came out of reflections on questions introduced in Manuel DeLanda’s book “War in the Age of Intelligent Machines” (DeLanda, 2001). In particular the following question was posed: if humans became extinct now, how would an autonomous Artificial Intelligence (AI) interpret leftover human cultural remnants, images, and artifacts devoid of a human context or operator? How would a lone AI view its own technical evolution? This idea has been explored through artistic research and the results will be presented alongside examples of other works that stemmed from the creative interactions with these algorithms. There will also be a reflection on the technical and more poetic aspects of working with GANs and ML for creative production. Through the use of internet-based data-sets sourced from freely available online material, we have been experimenting with how a GAN will interpret particular aspects of material and visual culture, and possibly reveal elements that are not completely obvious to a human observer through its generated output. We have observed that certain patterns can be seen in the GAN- generated material and can offer insights into the data-sets but also into the normally black box processes within the GAN itself.

This talk will move through the different trials and results that developed through this project. It will start with the first experiments of training an image-based StyleGAN2 model on thousands of images of car crashes and random stock images to produce distorted hybrid images. We will also speak about the development of the 3D voxel-based GAN created by our team that we trained on over 2,000 3D models of hand tools.