Thursday 2 February 2012

Existence & Computational Creativity

In this paper we are going to be looking at computational models of creativity. Computational models are generally applied in two different projects, in understanding human creativity and in producing machine creativity.[1] The idea that computational models can take us some way towards explaining human artistic creativity is not uncontroversial. In this paper I will be looking at reasons to think that computational models are inadequate to map what is essential to artistic creativity but that they have important explanatory value nonetheless.

Before we can begin a definition of creativity will be necessary. There is broad consensus in the field on two necessary conditions of creativity which is thus defined as a faculty for producing (1) original and (2) valuable things.[2] So with a rough idea of the minimal basis for creativity we can better see what forms creativity can take. Margaret Boden identifies three such modes; combinational, exploratory, and transformational.[3] Combinational creativity involves making original and valuable connections between things which haven’t been made before, as with metaphor or analogy. The exploratory and transformational forms require more elaboration.

In order to understand what exploratory and transformational creativity are we have to first know what a conceptual space is. Boden characterises a conceptual space as a “structured style of thought” which does not originate in the individual’s mind but is rather picked up from their cultural or extra-cultural environment.[4] To take a modern example, dubstep music focuses on syncopated rhythm structures, sub-bass frequencies, minimalism, little or no vocals, and largely minor key melodies. These sensibilities constitute the “conceptual space” within which the music is created. The creation of a dubstep track constitutes exploratory creativity, i.e. the creation of new things from within the conceptual space which exhibit its general sensibilities.

However, this conceptual space wasn’t simply “given” - it had to have come from somewhere. Around the late 90’s the then popular 2-step garage began to transform into “dark garage” and then into what became known as dubstep. The syncopated rhythm structures of 2-step were assumed and coupled with elements from dub music, resulting in a focus on sub-bass, lack of vocals, and minimalism (hence the name “dubstep”). This constitutes transformational creativity as the conceptual space itself is transformed opening up the possibility of creating music within a new space under new sensibilities, music which could not have been created before the transformation.

Boden’s claim that exploratory-transformational creativity (henceforth “ET-creativity”) is measured according to a conceptual space gives us a great way of evaluating a given non-combinational creative act. The above example shows that her model of ET-creativity can successfully be applied to at least some real world cases, and we can easily imagine others to which it would fit. Furthermore, combination and the exploration/transformation of conceptual spaces can be mapped by computational models, a fact which gives Boden reason to think that such models are able to help us explain creativity. Many attempts have been made to create computer programs which help us understand creativity and a few are even able to demonstrate some form of transformational creativity.[5]

In the following I will present reasons to think that computational models may not be able to help us adequately understand creativity and certainly not to emulate it, but I do not mean to claim that computer programs can’t model any aspect of creativity at all. Insofar as certain parts of the creative process involve deduction, the use of heuristics, and other computable operations it would be a mistake to insist that computer models can’t map these aspects of creativity. The claim is simply that what is essential to artistic creativity may not exhibit a formal computational structure.

Now, an additional condition of creativity offered by Berys Gaut is that a creative act must demonstrate understanding. The intuition is that if an act did not emerge from deliberate intentionality then that act cannot be called creative. Gaut’s example is of tectonic forces creating diamonds, a case which clearly satisfies the originality and value clauses but which we’re reluctant to call a creative act.[6] We hesitate to call this creative because it doesn’t involve agency, but as Gaut points out not just any old type of agency will do. If we were to spill paints on a canvas accidentally and it resulted in a nice looking piece we still wouldn’t think of it as a creative act as it did not involve the deliberate application of skill.[7] If it turns out that computer programs cannot understand the ends to which they are set, on Gaut’s definition they cannot be said to be creative. But this only suggests that computer programs can’t truly emulate artistic creativity - the question remains as to whether they can at least help us explain creativity in some way.

An account of human intelligence offered by Hubert Dreyfus claims that our ability to understand things depends upon our concernful, active way of existing.[8] Dreyfus became infamous in 1964 for predicting that computer programs attempting to create machine intelligence would run into trouble when they tried to represent more intuitive forms of understanding.[9] Around the mid 1970’s difficulties encountered by information processing computer models led AI researchers to believe that our cognitive information processing capacities in fact operate upon such an intuitive background understanding, a predicament which became known as the “common sense knowledge problem".[10]

The goal, then, was to represent this common sense understanding in a computer model. When the task of representing the whole of common sense knowledge began to look overbearing, AI researchers created something called a “micro-world” to make the task more manageable.[11] A micro-world is an isolated aspect of human discourse, like being able to understand a story, follow instructions for moving blocks, or indeed be creative. The assumption behind micro-worlds is that fewer common sense primitives will be required in order to generate results, given that we’re only emulating one aspect of human discourse at a time. Marvin Minsky and Seymour Papert consider the common sense primitives a computational model would need to understand a simple natural language statement offering a trade. The list includes temporal categories (to grasp that you have an item presently and in future may have another in exchange for it), values (to grasp that the item you have may be better, worse, or equal in value to the item exchanged), and so on.[12]

The idea of a micro-world proceeds from a controversial philosophical assumption – that you can create human intelligence by working on each human capacity separately and that in doing so you will eventually converge on a set of computable primitives which account for all aspects of common sense.[13] Furthermore, the claim that humans represent common sense understanding by means of cognition is also a controversial assumption and stands in need of justification.

Boden’s insistence upon exploring or transforming a conceptual space reveals her emphasis on our “higher” cognitive capacities, but in order to behold a concept one must first have initially grasped something as what it is. Martin Heidegger, the source of much of Dreyfus’ claim, argued that lived experience must necessarily precede the having of knowledge.[14] What this implies is that having concepts must first be preceded by a concernful engagement with the world in order that a given phenomenon can be understood as something and thus rendered as a concept a posteriori. Once a mountain is grasped as a mountain and gold grasped as gold we can imagine Hume’s golden mountain. So how do we understand our sense impressions and secure our perception of a thing as a thing?

One of Kant’s lasting contributions to philosophy was in raising the issue that direct intelligibility of the sensory manifold is problematic and that we stand in need of an explanation as to how sense impressions are intelligible to us at all. Kant himself sourced intelligibility in his categories of the understanding, thereby subjectivising the source of intelligibility and emphasising the role of separate individuals in its accomplishment. When this problem landed on Heidegger’s desk he sourced intelligibility in what he called “the world”. The world in Heideggerian terminology denotes a publically shared nexus of concernful activity signifying equipment, people, the natural world, etc. which hangs together through the pursuit of possibilities.[15]

Only within “the world” can something demonstrate itself as what it is. To make this clear consider again the notion of trade then imagine two instances of a person handing something to someone. As far as raw sense impressions are concerned the situations could be identical, but imagine that in one a debt is being repaid and in another an item is being given to someone as a gift. The situations mean different things with respect to the purposive involvements of each person and are in fact only discriminable through common sense familiarity with such customs (the sense impressions themselves can’t get us this far as both situations look the same). Heidegger would insist that we can only understand the difference between the two cases by being a purposive being oneself and thereby being able to understand the different ends involved.[16] Both are possible ends human beings can pursue and so bringing a pre-conceptual grasp of such ends to bear on our sense experience actually renders it intelligible.

If the foregoing is correct, our everyday common sense understanding of the world must precede cognition. If there wasn’t this original phenomenological encounter, along with the way of being which renders things initially intelligible, our conceptual information processing capacities would have nothing on which to operate. Our problem is that a human being has to program computational models[17] with a conceptual representation of this kind of background understanding in order to get them to perform in creative micro-worlds. If Dreyfus is right, human beings do not have to be programmed in this way but rather already have such common sense understanding as the simple result of our being what we are.

Computer programs model our capacities separately when what seems to be most interesting about them is their fundamental inseparability. Dreyfus’ point is that a given feature of human existence like creativity can’t be considered in isolation from the rest of our concernful activity, customs, and practices. Our capacities do not stand as a composite of isolated skills but are rather related to one another as aspects of the same holistic phenomenon rendered possible by the active, engaged kind of being[18] as which human beings exist. Computational creativity therefore suffers from the common sense knowledge problem as much as any micro-world.

To make this explicit consider Picasso’s Guernica, a painting about the German bombing of Guernica during the Spanish Civil War. In order to create such a work in a deliberate, intentional manner Picasso must first have been able to grasp the significance of such an historic event. One would have to have this kind of common sense understanding for war to demonstrate its character as horrifying or lamentable or even triumphal, as the case may be. Similarly to grasp Marcel Duchamp’s Fountain one needs to be familiar with the customs and practices of the art world to appreciate how it stands out as a curiosity (again, raw sense impressions cannot convey this – it looks like any other urinal).

Boden actually anticipates that creativity is in part a socially integrated affair but nevertheless insists that computer models can tell us about how creativity emerges in individuals.[19] However, it is the initial having of a conceptual space which is the real mystery of creativity and not the calculative procedures which one may follow whilst already in such a space. The “magic” of creativity is still being performed by human beings with computational models simply able to follow out any residual deductive or heuristic procedures. Even the impressive transformational models Boden mentions, such as EURISKO and AM, presuppose a large degree of human involvement. Douglas Lenat himself admits that concepts given in the LISP programming language and even human guidance in problem solving are partly responsible for their success.[20] So long as computational models of creativity have their common sense background represented ‘conceptually’ in isolated micro-worlds they cannot get us to the heart of what is most curious about artistic creativity - that it emerges from an inseparable connection with the wider phenomenon of human existence.

Boden has said that the aim of computational models is to help us understand creativity, even if that means by falsification.[22] To this extent the successes and failures of computational models do in fact help us to explain creativity. Even if they cannot at present represent what is most significant about artistic creation these models nevertheless allow us to see whether or not creativity is computational in part or as a whole and so in this respect they have significant explanatory value.

- Jordan Adshead (January 2012)

Bibliography

Boden, M. A. (1999). Computer Models of Creativity. In R. J. Sternberg, Handbook of Creativity (pp. 351-372). Cambridge University Press.

Boden, M. A. (2004). The Creative Mind: Myths and Mechanisms (Second Edition). Routledge.

Dreyfus, H. (1992). What Computers Still Can't Do. The MIT Press.

Dreyfus, H., & Dreyfus, S. (1987). Mind Over Machine. The Free Press.

Gaut, B. (2010). The Philosophy of Creativity. Philosophy Compass, 5, (12), 1034-1046.

Heidegger, M. (2010). Being and Time. SUNY.

Lenat, D., & Brown, J. S. (1983). Why AM and EURISKO Appear to Work. Proceedings of AAAI, (pp. 236-240).



[1] Boden (2004), p. 1

[2] Gaut (2010), p. 1040

[3] Boden (2004), p. 3

[4] Ibid. p. 4

[5] Boden (1999), pp. 365-368

[6] Gaut (2010), p. 1040

[7] Ibid, p. 1040

[8] Dreyfus (1992), pp. 285-305

[9] Dreyfus & Dreyfus (1988), p. 7

[10] Ibid, p. 68

[11] Ibid, p. 72

[12] Ibid, p. 74

[13] Ibid, p. 74

[14] Heidegger (2010), pp. 59-62

[15] It’s important to make explicit that Heidegger doesn’t imagine the world as an exhaustive list of all objectively present beings as one might ordinarily imagine, but rather imagines activities, moods, capacities, and so on as a part of the world too, things which cannot be understood with regard to objective presence.

[16] Ibid, pp. 81-87

[17] Boden (1999), p. 353

[18] Being in the active sense, i.e. be-ing, like mov-ing, or walk-ing. Not to be confused with being in the sense of a being.

[19] Ibid, p. 352

[20] Lenat & Brown (1983), pp. 236-240

[21] Boden (1999), p. 353

No comments:

Post a Comment