Fall of the House of Usher II
This is the training set that was used to create part of 'Fall of the House of Usher I'. The training sets, or images that are given to machine learning algorithms as input, provide the knowledge and are central to the eventual image output, yet rarely discussed in an artistic context. Most of these training sets that are readily available are compiled by researchers, often using mechanical turks using various methodologies. Because people are always involved at some point in either the source content or in the process, they inevitably come to enshrine cultural or social attitudes, otherwise known as dataset bias. I am aware that the control that comes to me in this process really comes from what I do with the dataset. As someone who is interested in the hidden and the forgotten, it makes me uncomfortable to use someone else’s dataset without properly exploring it. Small batch machine learning programmes (such as pix2pix) require a much smaller training sets (often only in the hundreds), which means it is possible to explore and think through these ideas. By creating my own dataset, it forces me to examine each image and inverts the usual process for creating this type of dataset. This particular training set is a series of paintings that I have made of imagery created by machine learning for part of 'Fall of the House of Usher I' (the piece consists of an animation made from three separate neural nets – the first trained on drawings of the original frames, the second on drawings made of the results of the first net, and the third on drawings made of the results of the second). Painting these was incredibly difficult - the logic of the world is there but not quite there (shadows are not quite right, twists in fabric do not hang as expected). Style in drawing always evolves but because I drew in this style for so long, it has changed how I now draw and paint a picture. I now start to add in artefacts and my line has changed. Working in this way, I am not interested in programming a machine to draw like a human or in producing a drawing that does not acknowledge its origin. The shiny, robotic quality of much digital art appears to neuter the messiness of the world. I am interested, however, in the opposite approach: how to use a medium that is cold, sterile and algorithmic to maintain and accentuate a sense of human touch. By laboriously creating by hand datasets based on the original film and then by processing and reproducing these by using specially modified algorithms, I have created a system of loops and feedbacks that use and enhance machine learning as an integral part of the work’s material and process.