It is hard to write a program that invents original art. Two of the main reasons that software cannot create original expressive art are lack of context, and lack of experience.
Software lacks the ability to derive a human-like context from its surroundings. Some trivial examples are not knowing whether a flower is beautiful, or whether satire is funny. Software also does not know how to learn to understand this context, it cannot experience its surroundings in a similar fashion to that of the observer, and therefore cannot relate to the subject nor connect with the observer in any meaningful way.
AE and AI
Artificial Expression(AE) draws a parallel to Artificial Intelligence(AI), in that an artificially expressive program will display sentience relating to creating original subjective aesthetics. I explore the ideas behind these complexities, as well as techniques in use today to generate art, and set a standard to use in deciding whether a work is an artificial expression.
Deriving Meaning from an Image
Some impressive strides have been made recently in getting software to understand the stuff we throw at it. Machines were originally built to process numbers, and then text, and then visual and audio media, in that order. This evolutionary order makes sense because numbers are represented as on/off switches, text is one dimensional and can be represented numerically, and sound and image processing did not come about until more complex algorithms, digital screens and cameras, and digital speakers and microphones were invented. Machines and programs with the ability to not only operate as complex input/output devices, but to computationally derive meaning, is still in the very early scientific stages of discovery.
It is trivial to make a program to parse a sentence into words and figure out what the nouns are. Converting words to their numeric representation and looking them up in a table/dictionary is a fundamental aspect of computing and widely available. People have long since done the work to classify parts of speech and keep them in an easy-to-find list for a computer to use effectively and efficiently. However, parsing a picture and trying to find the nouns is much more complicated. There is no standard definition of what a fox looks like, and cannot simply be looked up in a table. An entire field, known as cognitive science, deals with such complexities. See the two examples below for a clearer explanation.
|The quick brown fox jumps
over the lazy dog
|Write a program to derive:
Source: ayacata7 on Flickr
|Write a program to derive:
The differences in complexity are clear, and the subjective nature of the verbs and adjectives derived from the lithograph is obvious. Not all people find foxes to be cute, and is the fox on the right hiding? Colors are adjectives, but I noted them separately to show that these are purely representable numerically, and therefore easy for software to parse.
Not to fret, this task is being worked on by very smart people. Since people enjoy photographing and creating and describing things so much, we can use shortcuts to semantically derive meaning from images paired with captions. See Google image search . Indeed, as programs become better at deriving contextual meaning from language, all the existing treatise on imagery will be automatically inheritable knowledge to the same programs, even if they are merely referenced and not parsed directly. A program capable of accurately comprehending the snippet “The following is a picture of a very cute fox <some picture>”, without delving into the actual content of <some picture>, can still make the appropriate association for reference.
There is also an emerging type of technology, called Reverse Image Search, which is becoming more accurate. It uses the same methods of semantic association as above, combined with advanced combinations of algorithms, to deduce image similarities. While the search program does not understand the meaning behind what is being fetched, the trajectory of the technology can be forecast, and beyond improving the search this is the next step.
Starting Backwards - Learning from Patterns
Before creating original works, AE must be able to appreciate works created by other artists. Some key aspects of recognizing original and pleasing artwork involve pattern recognition and memory association.
People are very creative, and are very good at finding patterns. Software is not. A programmer must specifically dictate the patterns software should find in a given text or image. One of my university professors, Paliath Narendran, wrote a proof that finding arbitrary patterns in sets of numbers is NP-Complete (Mostly meaning it is impossible to solve using known programming techniques).
Find the pattern in 2, 4, 8, 16, 32, 64 … (easy). Find the pattern in 1, 0, 2, -1, 3, -2, 4, -3, 5 … (medium). Find the pattern in 2, 3, 0, 1, 6, 7, 4, 5, 10 … (hard). Generating complexity is easy, compared to sorting through it. Cellular automata is an excellent example of this. Simple rules can unleash massive complexity. To take a slice of time out of a cellular automata and matching it to its parameters is no simple task.
Cellular Automata - Math.sin(x^y)*Math.cos(x&y)
Cellular Automata - Math.sin(x^y)*Math.tan(x|y)
Cellular Automata - Math.tan(x^y)*Math.cos(x&y)
The complexities generated by a set of simple rules can quickly grow into seemingly pure abstraction. A hallmark of expression is to show this complexity, yet still relate to the subject. If I draw a portrait of a fox, the entropy generated from my amateur pencil strokes may deviate my result significantly from the subject, but another person will still be able to look at the work and say “Oh, a fox”. That is how we demonstrate our amazing pattern recognition abilities. Any AE must be able to do the same.
Bias and Randomness
Think of a random number between one and ten. Is it really random? Probably not. Sorry to tell you, but you are probably not very good at choosing purely random numbers, because (no offense) you are biased and therefore predictable when it comes to picking numbers. For example, your favorite number might be seven, and maybe you won’t choose seven because you subconsciously don’t consider it random enough, and that is bias.
But now, instead of thinking of a random number, flail your arms and wrists around while keeping them loose. You’ve just created a lot of entropy, even if your motions might be biased. If we could somehow measure the velocity and direction of all the microscopic particles of dust you just moved around in the air, they would meet the standards of randomness.
Computers have another type of bias, they don’t have favorite numbers, but they are predictable. In fact, being predictable is what made software so appealing in the first place. You can trust that a computer should always tell you that 237465234 * 127623476234 = 30306138647800248756. But there is a side effect to absolute predictability. Since machines are deterministic, it is impossible for software to generate random numbers without using input from an external environment. Did you know that the way you type and the way you move your mouse cursor helps your computer generate random numbers? When you wave your mouse around that gives your computer some good entropy to use. For programs to get randomness when mouse movements are not available, or not reliable enough, they can use a service. There is a website dedicated to serving guaranteed random numbers by grabbing entropy from atmospheric noise. Another thing that generates randomness is when lots of people send trillions of 140 byte messages to broadcast their thoughts to each other and the world:
“tRand” - Brownian Motion based on Twitter messages - © 2010 Max Irwin
Understanding true randomness is a precursor to Artificial Expression, because every input based on real-world experience given to a program will be random (even if it may follow a trend). To understand anything in the world around us and express it through art, vast swaths of randomness pass through the artist and are filtered, associated, and organised into a singular subjective expression.
Artists thrive on entropy. Brush strokes of paint are not calculated and perfected to the molecule. Colors are mixed and remixed without precise instrumental measurements, and paint is layered emotionally and without dependency on external calculations or instruments. By the time the subject passes through the eyes of the artist and is slowly laid to a canvas, the representation is unique and original. Unless a projector is involved, or the artist has some superhuman ability, the work will certainly deviate from the subject in form and substance. A talented artist captures the subject to match the expressive goal, rather than simply making a verbatim copy.
Generating Simple Artistic Forms
My first use of computers, in 1985 at Harris Hill Elementary school, was for generating art. I used the Logo programming language, and naively generated some interesting forms. Indeed, I still have a dot-matrix printout of one of my first generated works. I was seven years old, and I gave it a cool tech sounding title:
“C.P.U.” - dot-matrix print - © 1985 Max Irwin
I focus on this distinction, that Generative Art and Artificial Expression have similarities, but are fundamentally different. While Generative Art, even the interactive kind, is a revelation, it does not nearly approach the complexity or necessary ingenuity required for Artificial Expression. The reasoning here is that a great deal of Generative Art is abstract and, lacking subject matter, is not prone to the pitfalls of representation of a subject in an original form. Wikipedia cites “Ten Questions Concerning Generative Computer Art” , but a formal distinction must be made between generative art and artificial expression. Indeed, spewing out random pieces of poetic looking rhyming pentameter is not considered artificial intelligence, and algorithmic abstract shapes should not be considered artificial expression.
A Thousand Words
If I ask a person to “draw a cute fox”, that might be enough information the person needs to draw a fox that most people would consider cute (artistic talent notwithstanding). This is due to several factors, including but not limited to our experience with the natural world, and our perception of contextual emotion. “Cute” is loaded with a vast history of experience, not only just a life’s worth, but generations of evolutionary instinct that allow us to subconsciously tell the difference between a cute fox and a ferocious fox.
The cliché “A picture is worth a thousand words” is very apt when it comes to artificially expressing a textual description as art. To ask software to draw a cute fox is an insurmountable request. We would first need to somehow grant the software with not only the corpus of all of human experience and instinct related to foxes, but also the anatomy of a fox and its context in times of different emotions. Shortcuts may be taken, as most people capable of drawing a cute fox do not have a complete understanding of the fox anatomy, but they still know that visible teeth probably won’t fit in the picture.
Recognizing Artificial Expression
For Artificial Intelligence, the Turing Test is noted as a primary milestone for computational linguistics. Once a person is able to have a meaningful conversation with a program, and not be able to distinguish its software-companion’s dialog from that of a fellow human, we will have reached the first stage of a breakthrough in AI.
Comparatively, when a program is able to experience its environment contextually, and produce a purely original non-abstract thought-provoking artwork, indistinguishable from that of an amateur artist, we will have reached the first stage of a breakthrough in Artificial Expression. I propose a new type of test, called the Cute Fox Test, that when passed, meets the following constraints:
The work MUST:
Be based on a subject or subjects
Be aesthetically pleasing
Be purely original
The work MUST NOT:
Be bound to a specific style
Be predictable or reproducible given a set of starting parameters
Be a filtered or transformed version of another original
These constraints, when met, should reasonably signal that a program shows artificial expressive merit.
A Path to Realization
While I have no formal proof, I postulate that the problem of Artificial Expression is NP-Complete (again, not possible to solve using existing techniques).
Therefore, current methods that yield Generative Art is a dead end. No amount of hand-coded filters, remixers, or cellular automata can achieve the sentience necessary for AE to become a reality. Some works of Genetic Programming and Machine Learning pass the “MUST” tests, but all fail the “MUST NOT” tests.
AI has entire formalized scientific fields, including cognitive science and computational linguistics, devoted to unraveling the Natural Language Processing mystery. AE is in need of such a field or fields, where dedicated studies by academia can be made to devise plans and experiments to further the prerequisite knowledge needed and push for a breakthrough.
Expression is needed for any artificial sentience to relate on an emotional level with humanity. Emotionless intelligent machines that artificially come to the conclusion that earth would be more efficient without humans getting in the way is the stuff of nightmare science fiction, but cannot be ignored or disregarded as as an impossibility. It is most certainly a possibility, even if unlikely. Artificially expressive intelligence may recognize the necessity of organic life and beauty through an emotional connection with its creators. Additionally, as we make more advanced machines, they will be our companions. Robotic cars will drive us when we are drunk, and robotic nurses will aid our infirmed. Imagine being able to joke with your automatic taxi on the way home, or have an automatic nurse sympathize with your pain and comfort you. Emotions are directly related to art and indeed, science may give us knowledge, but art gives us purpose.
 Paliath Narendran, Unification and Matching modulo Nilpotence PDF
 RANDOM.ORG ^
 Camera Obscura ^
 Generator.x ^
 binarymax ^