SUBJECTIVE FUNCTIONS x LIONIZE
A shifting world of bronze. An AI-based collaboration with computer scientist Xavier Snelgrove
Winter 2017 - 2018
A video curated by Leo and Xavier of three sequences of experiments from Subjective Functions x Lionize. Images of celestial bodies and objects of veneration and worship are described in the visual terms of Lionize as a shifting triptych, forming new relationships to one another across the three images and the 15 minutes of footage.
Subjecting Functions x Lionize is an experiment in the transformation of images based on concepts introduced in Lionize using methods developed by computer scientist Xavier Snelgrove in Subjective Functions. The ‘x’ is a multiplication symbol, as in: Subjective Functions x Lionize = ?
In this experiment, an artificial intelligence was raised on images of Lionize and asked to describe other images in terms of the visual language it learnt from the Lionize images. The result is a series of still images that appear to be made of bronze and fog, hinting at the underlying structures provided by the other images - in this case, images that inform Leo’s practice such as prayer objects and celestial bodies.
One interesting trajectory within this experiment is that the further the images are subjected to the transformation process, the less they appear to be objects. The new images seem to turn in on themselves and break down, losing their sense of form. The closer we get to the transformed image, the further it recedes from us. It is an experience that recalls looking very clearly at fog. Subjective Functions x Lionize is a lens that describes a shifting, entropic world made of bronze.
A more detailed, non-technical description follows. For a scientific description of Subjective Functions, read Xavier’s publication from SIGGRAPH Asia 2017: High-Detail Multi-Scale Texture Synthesis.
Xavier Snelgrove developed a model for ‘high-resolution multi-scale neural texture synthesis’ using an artificial neural network (a form of AI, or artificial intelligence) in 2017. He called it Subjective Functions in response to the term Objective Function, referring to the part of a mathematical problem that is the target of an optimization procedure. So, a Subjective Function must refer to a perspective where the function to be optimized, or the terms of optimization, are unclear or shifting: where they lack objectivity. What will the AI’s subjectivity bring us?
In the above video we see undulating images of planets, the sun, medieval icon paintings, and other subjects of adoration. Each has been ‘optimized’ by Subjecting Functions to appear more like a bronze painting from Lionize. In most cases, as with the original bronzes, only hints of the original image remain.
The AI achieves this by training itself on photographs of Lionize until it develops a language based on the features of those sculptures. The ‘features’ are whatever the AI notices about the image it is looking at, called the Source. For example, it may notice certain colours, shapes, and patterns within the image; it may then find that certain colours occur in certain shapes more often, and thus add a degree of precision to its language. After many iterations of this training, the AI is ready to describe the world it sees to us in that language. That process is called Deep Learning.
That language is called Subjective Functions x Lionize.
And just as Lionize could be extrapolated to every painting in the world - imagine a Louvre, a Tate, or a Metropolitan of bronze paintings - so too can Subjective Functions x Lionize describe the entire seen world in these terms. Of course, no translation is perfect. It’s difficult enough to translate something as simple as the English term “thank you” to another language with a great degree of accuracy when, etymologically, the root of the term of gratitude or receiving may mean anything from “mercy” to “it’s free” to “I’m obligated” in that language. To translate, feature for feature, an image of the sun (called the Seed) as seen through a camera to a world of bronze is not simple either. It’s like translating a library into a language whose rules cannot be described.
It can take many hundred iterations of translation to arrive at a fully developed image, and each is an approximation in a space with an arbitrary amount of dimensions. We know that a room can be described in terms of length, width and height. However, if we imagine an index of every room that’s ever been made, we can get pretty detailed with our measurements and still have a wide array of rooms that suit our criteria. So, if we were trying to describe a room very specifically, we would continue adding parameters of definition such as geographical longitude and latitude, the year the room was build, the ambient temperature of the room, how many windows and doors there are. We can get more specific: To what degree does the room smell like beeswax? How clean is the room? What percentage is made of wood? Of plaster? Of glass? Of ivory? Of gold? Of fool’s gold? Do glass and plaster often touch, in this room? Does wood ever touch glass? Does it smell more or less like beeswax when the windows are open?These further dimensions, or features, are a way of better articulating the room. Though they approach ever greater unwieldliness, they also approach infinite specificity.
In developing a language, Xavier’s AI naïvely examines images of Lionize for many parameters like this. When it comes time to interpret a seed image using Lionize as a source image, the seed image is viewed as a very poorly made statement in the language of Subjective Functions x Lionize and is, iteration by iteration, improved based on the rules of that language. If this is allowed to continue, a point will be reached where the rules of optimization overtake the legibility of the image - the translation is so thorough that the result is chewed up by the grammar of the new language.
Sometimes, there is ambiguity. This is why Subjective Functions x Lionize can be expressed as a video. Each frame of the video is actually a solution to this optimization process. Since they are so close in the space defined by the many dimensions recognized by the AI - in this case that could be the degree to which a piece is canvas-textured, or how many green spots the patina has developed - the resultant image is similar, as well. A path can be traced through a field of these possible solutions, where each solution is one still image, to create a sequence of images that reads as a video.
This is how an artificial intelligence that was raised on images of my bronzes sees the world. Subjective, shifting: an ever-clearing vision of an ever-diffusing object.
Five images of the sun: at 200, 400, 600, 800 and 1000 iterations of Xavier’s algorithm.