What Does it Mean to Create Something Valuable — and Can a Machine Do it For You?
As Artificial Intelligence becomes more and more proficient, will it become capable of replicating one of the most human capabilities?
Much of what we appreciate in art is subjective and intangible; our reaction to a piece of work is often a personal and emotional one. But a lot of what guides our positive responses to art — particularly in areas like music and in visual media — is hard-coded in nature. How much of our enjoyment of a melody is simple pattern-recognition? AI may be able to identify what we really appreciate in paintings using eye-tracking technology. Do I prefer one photo over another just because it follows certain rules of composition? And if the basis of much of our appreciation is in numbers and logic, can powerful machines and sophisticated algorithms predict what will stimulate a response and construct it themselves?
Artificial Intelligence has already demonstrated its ability to reliably take on repetitive, predictable functions that can streamline and automate what would have been time-consuming tasks. AI thrives in a narrow, well-defined space with set rules and sequences, and its ability to assess data, find patterns and make predictions massively exceeds human capabilities. Natural language processing and text analytics used to develop voice assistants like Siri and Alexa are surpassingly more intuitive and reliable than they were even a few years ago, while advertising copy written by AI is now so good that customers actually prefer it to human-written material. The capabilities of deep learning algorithms have leapt forward in recent years (driven in a large part by investments by Google and its Deepmind lab) and these improvements promise to drive technical advances to even greater heights. Is it only a matter of time before AI can make the jump from understanding language and making predictions to formalising human aesthetics and creating original, meaningful novelties?
Rhythm depends on arithmetic, harmony draws from basic numerical relationships, and the development of musical themes reflects the world of symmetry and geometry” –Marcus Du Sautoy
Music theory is one branch of the arts that relies on patterns and repetition. Early Greek writings specify the mathematical proportions involved in tuning systems, for example, while octaves, scales and semi-tones are the regular and predictable building blocks of music itself. These conventions are uniformly observed across musical styles and cultures — even the music of Balinese gamelan singers that utilise divisions of notes that do not exist in Western Classical music rely on the mathematical measurement of time and frequency. Virtually all recognisable music has a relationship with geometry. This may suggest that music is intrinsic to nature (nature itself may be described as “amazingly mathematical”), and not a social or cultural construct that has grown throughout a civilisation over time.
Humans are natural pattern recognisers, a skill which is a fundamental component of what we regard as intelligence. Our enjoyment and appreciation of music may be a result of our evolutionary development of recognising regular patterns in the natural world — or it may simply be that some things sound good next to other things and we don’t really know why.
What we do know for sure is that the patterns that govern these conventions are totally predictable. I could (in theory) write a rudimentary melody using standard scales and chord progressions that a mountain-dwelling hermit who had never heard music before should recognise as aurally pleasing, even though I have effectively no musical training and I’m terrible at the piano. There may be some cultural inconsistency (much like the Balinese gamelan might seem overwhelming to someone who has only been exposed to Western music), but the hermit should still, after a time, begin to appreciate the innate pleasure of the repeating patterns.
If these conventions are so predictable, is there anything stopping a computer algorithm from implementing them and writing music as well as a skilled composer could? A deep learning algorithm could theoretically have access to the entirety of recorded human musical output, making it far better educated than the world’s foremost musical expert and arming it with the knowledge of how musicians have been successful in the past. But surely it takes something more than vast knowledge to create something unique.
One of the most mysterious and poorly understood human thought processes is creativity — a hugely sophisticated capability that has shaped humanity and the world it has created. It is the driver of all facets of human accomplishment; the mover that allows us to improve and develop as a society over generations. It has allowed us to pursue scientific, mathematical and philosophical learning and knowledge.
Animal species are (broadly) incapable of thinking creatively (while many species are capable of innovation, they cannot share information and build on these insights), and, heretofore, so are machines (or so the received wisdom would indicate). With massive strides in processing power over the last number of years, however, and improvements in machine learning and deep learning algorithms, computers are coming closer and closer to replicating the intangible inventiveness of creativity.
Creativity is understood in fields such as computer science and linguistics as the ability to bring into existence non-trivial novelties; i.e. that which is original and meaningful. Creativity in humanity is fostered by universal generators — systems and structures that have been developed through innovation and improvement that allow for the creation of persistent novelties (persistent in that they prove themselves to be useful over time).
Language, for example, is a universal generator that allows us to formulate abstract ideas into an understandable basis for communication. Any person who understands the grammatical rules of a language can create an entirely original sentence, or indeed a poem or a novel. If they don’t follow the rules, they are just speaking gibberish. The decimal system of mathematical numerals theoretically allows us to count to infinity — and music theory enables us to write a melody, harmony or symphony.
These universal systems were originally developed to solve particular and specific problems, such as basic communication amongst early humans, or mathematicians who required representations of larger numbers to complete their equations. But because these systems and structures allow for reach, that is, having effects and implications beyond what was originally envisaged, they become universal, and as such are environments that facilitate and foster creative conjecture.
Conjecture is the first-mover, that spontaneous spark of inspiration that launches an idea from the void into reality. When evaluating art, we associate value with the capability of a work to stimulate an emotional response. If a novelty fulfils the criteria of persistence, but fails to demonstrate value, the novelty remains an example of creativity, but it is not an example of artistic, or true, creativity. This distinction is where humans maintain a marked edge over machines, which are incapable of emotional responses. But who knows where future developments will take us?
How much of song-writing is down to pure inspiration, above gathering what we have experienced in the past and re-organising it into a new form? Musicians borrow openly and liberally from other sources (DJ’s and producers create music entirely made of samples recorded by other artists, for example). But ideas have always been reused, borrowed and re-worked — that is how society grows and develops. Any new idea is ontologically composed of older ideas — no thought occurs in a vacuum. We don’t create anything in isolation; rather, creativity is, at its core, a collaborative exercise that takes inspiration from a wider community. Our creative output is informed, shaped and driven by what we consume.
Artificial neural networks today can be trained to recognise music and generate it themselves. The algorithm that creates this music is basing the entirety of its output on music written by others, i.e. whatever examples of music its creators have decided to feed it. This kind of neural network is always bonded to its training data and is therefore incapable of constructing a worldview, or drawing from past experiences, outside of this relatively narrow data set, as humans can. The opportunity for true insight and creativity in this scenario is limited, and the resulting music is accordingly pleasant to the ear without being ground-breakingly original. But as technology becomes more sophisticated, we may see artificial neural networks utilise much broader training data to develop infinitely more interesting outputs.
The only original input from the machine when it generates music is the random experiments it makes while formulating and constructing it. Is this any different than the random and unpredictable sparks of inspiration that humans call ideas? The intentionality of the artistic direction is a differentiator between humans and machines, but given enough scope and time, the results should eventually resemble each other. The challenge for a human musician is to use their informed judgement to recognise which ideas are actually any good (i.e. have value and have the potential to persist), while machines that lack critical functions rely on humans to pass judgement on their efforts (although research in Generative Adversarial Networks could facilitate exciting advances in the future).
Any given output of such a machine is one example out of an infinity of possibilities. This in itself is a criticism of such work specifically as art; after all, if this is just one example out of an infinite series of potential results, what is interesting about it? Arguably, the method of its construction negates its value as art. Generative art created by autonomous systems that are currently being exhibited in galleries across the world, or the generative music of Brian Eno, while being aesthetically pleasing, could be described as conceptually boring. Such criticism may be legitimate but is predicated on a particular understanding of the nature of creativity. Alan Turing posited that a computers method of working out consequences and results from data was no different from how a human mind comes to conclusions. Machines, therefore, are just as likely to surprise us as humans are. This reflects Turing’s background as a computer scientist and stands in contrast to the view of creativity in the artistic tradition. It is a matter of personal viewpoint whether one regards the artefact or the process used to create it as more valuable.
Many companies are using machine learning techniques, coupled with massive data-sets to generate totally original music that can be adapted by artists and music professionals to whatever use they see fit. As we can see, modern systems are powerful enough to replicate examples of universal generators in an artificial, digital environment. This is true not only in the music world — AI can use massive computational power to develop scenarios and explanations through random trial-and-error that humans cannot begin to contemplate. Most of these results can be expected to be nonsensical or trivial, but with enough processing power, one will inevitably begin to find genuinely valuable solutions amongst the detritus. In the place of true human conjecture, this process might be said to produce brand new information that could be interpreted as creative were it to come from a human being. If the nature of reality itself is “amazingly mathematical”, the potential and reach of these systems are virtually limitless.
Music is true. An octave is a mathematical reality. So is a 5th. So is a major 7th chord. And I have the feeling that these have emotional meanings to us, not only because we’re taught that a major 7th is warm and fuzzy and a diminished is sort of threatening and dark, but also because they actually do have these meanings. It’s almost like it’s a language that’s not a matter of our choosing. It’s a truth. The laws of physics apply to music, and music follows that. So it really lifts us out of this subjective, opinionated human position and drops us into the cosmic picture just like that. — James Taylor, Performing Songwriter May 2002
Machine learning algorithms can, over time, develop an understanding of which newly-generated novelties are truly persistent and useful. In this regard, AI can be said to possess some of the attributes of a creative mind — that is, to construct original and meaningful solutions. Machines are not spontaneous; they do exactly what they are told to do (even though this may not always be exactly what the programmer thought they were told to do). But machines can - when instructed to learn - develop insights and cognisance of realities that humans cannot. This, perhaps, is the machines own, unique creative voice.
All views are my own and not shared by Oracle. Please feel free to connect with me on LinkedIn
