I once wrote about AI learning a “language” to communicate among themselves (and “encrypting” it so that outsiders or other systems can’t understand it). Recently, there has been a new development regarding language and AI. This special attempt seeks to decrease visible content that “AI can see” and turning it into symbols, similar to artworks, that another AI can mutually understand. Come to think of it, it is similar to Chinese characters or Egyptian hieroglyph.
We know that Chinese characters originate from Oracle Bone Script. They are characters written on animal leather or bones in the 2nd millennium BCE. This ancient character looks somewhat like drawings wherein the character for the sun is a circle with a dot at the center or the character for water is curving lines and drops of water on the sides. As time past, the letters evolve to become more systematic and substituting more abstract meanings. Later, it became the characters that we see and use today.
This type of character is called the Logographic system, meaning it uses characters in place of words or phrases. There is historical evidence that this system is part of languages in Africa, China, and Central America. However, a language solely relying on Logogram may not make lives any easier (except for languages with no speakers such as the Toki Pona that was constructed with only 120 words). Therefore, most languages rely on Logogram coupled with classifiers or other elements to construct dialogues to completely and meaningfully communicate.
So, what does this got anything to do with AI?
Tom White, a computational design professor at Victoria University of Wellington School of Design developed an AI that can classify objects and turn them into meaningful symbols. This resulted in abstract artworks and Tom White named the system the “Perception Engine”.
(The artworks are intentionally named to mimic Belgian artist, Magritte’s “This is not a pipe”.)
White began by questioning whether we can create abstract work from substantial sets of images. He is aware that, today, AI can precisely classify images and objects (such as distinguishing dogs from cats or determining how many people are in the photo). However, what he wants to know is that after AI has seen numerous images of a certain object, can they define the core or common essence of such object (such as an electric fan must have blades and a central motor). His system consisted of three parts, planning, drawing, and creative objective, all three systems work together.
The fundamental of the experiment is that after the system has created the image that represents the object, White tested the images by asking another AI to interpret them. The AI that decides (somewhat like audiences) were not involved in the creating process. (However, the interpretation was multiple choices, not open-ended questions. Similar to when you are asked to look at a picture and given choices to choose what you see).
The fact that only AI can precisely interpret the images (not so many humans can define what AI intended to create), the FastCompany magazine is calling Tow White’s work “Art by AI, for AI”. White’s artwork (and the AI. - Oh! or is it AI’s work in collaboration with White?) is for sale under the name The Treachery of ImageNet on BigCartel.