Being Kasimir Malevich: a reflection on computational creativity

Artemis Moroni, Dr.

Computer Vision and Robotics Group, CenPRA, Campinas, Brazil.

e-mail: Artemis.Moroni@cenpra.gov.br

Abstract

Recently, evolutionary art systems have been suggested to simulate some kind of creativity in computers. Here, a reflection on computational creativity is presented, mentioning classic and historical aspects and showing results of evolutionary art systems. Strong and weak features of these kind of systems are approached.

Introduction

How can computers have anything to do with creativity? The first person to denounce this apparent absurdity was Ada, Lady Lovelace, the friend and collaborator of Charles Babbage [1]. She realized that Babbage’s “Analytical Engine” – in essence, a design for a digital computer – could in principle “compose elaborate and scientific pieces of music of any degree of complexity or extent.” But she insisted that the creativity involved in any elaborate pieces of music emanating from the Analytical Engine would have to be credited not to the engine, but to the engineer. “The Analytical Engine has no pretensions whatever to originate anything. It can do [only] whatever we know how to order it to perform”, she said.

The Analytical Engine was never built but Babbage recognized that, in principle, his machine was capable of playing games such as checkers and chess by looking forward to possible alternative outcomes based on current potential moves. Turing [2] countered Lady Lovelace’s argument by equating it with a statement that a machine can never take us by surprise. But he noted that machines often act in unexpected ways because the entire current state or initial condition of the machine is generally unknown; therefore, an accurate prediction of all possible behavior of the mechanism is impossible.

But, until the present moment, it is not possible simply to say to a computer to “compose” or “paint”. Let us discard the interface problem, and suppose that someone may sit at the computer desk and simply type or select a command like “paint” or “compose”. One might say that even a person could ask “what” to paint or compose. How a medium person would answer to a general request for painting or composing? To be short, let us suppose that the required person would be able to create some compositions, trying to discover if he or she was pleasing. Can this human behavior be simulated in a computer?

1. Evolutionary Art Systems

A new generation of computer researchers is discovering that by using simulated evolution techniques it is easy to obtain novelty – often complex novelty – but it is correspondingly difficult to rein in the direction that novelty takes. These type of systems are commonly known as evolutionary systems. The loop in an evolutionary system is a rather simple one: generate, test and repeat. Basically, a bunch of things is made, tested according to some criteria and the ones that are better are kept.

The process is repeated by generating a new bunch of things – or population of any kind of objects - based on the old ones [3]. The loop continues for possibly many generations until the things that are being made are good enough according to the criteria being used. When an evolutionary system is applied to generate objects in the artistic domain, it is called evolutionary art system. Since it is difficult in the visual domain for example, to measure the aesthetic success of simulated objects or images automatically, the fitness is provided interactively by a human user based on visual perception [4]. Typically, in this kind of system, the generated frames are simultaneously presented in the computer screen for aesthetical evaluation.

In the musical domain, the temporal nature prevents the compressed, parallel presentation of individuals. First, musical objects cannot be presented in a compressed form without distorting them. The musical analogue to a reduced image would be a sped up musical sample. Even if the correct pitch is preserved, the tempo would be altered, which certainly changes the perception of a piece of music. The second problem is that multiple musical samples cannot be presented concurrently without obscuring the identity of each individual. The eye can focus on one image at a time, but the ear cannot isolate one melodic line from a randomly contrapuntal piece of music [5].

Even more ambitious is to replace in the process the human interaction by automatic critics trained using easy-to-collect examples. Baluja, Pomerleau, and Jochem [6], for instance, working in the visual domain, have trained a neural network to replace the human critic in an interactive image evolution system. Gibson and Byrne [7] suggested a very similar approach to generate very short musical fragments. The network “watches” the choices that a human user makes when selecting two-dimensional images from one generation to reproduce in the next generation, and over time learns to make the same kind of aesthetic evaluations as those made by a human user. When the trained network is put in place of the human critic in the evolutionary loop, interesting images can be evolved automatically.

2. Simulating Creation

But things are not simple. Let it be the famous Malevich painting, The Black Square on White Background [8]. This example was chosen simply because of its technical simplicity; it will not be considered the context and historical moment in which it was created. Kasimir Malevich was the artist who most embodied a new segment of modern art, the Suprematism. Malevich’s geometry had as starting point the straight line, the maximal elementary form that symbolised the human ascension over the chaos of Nature. According to him the square was the suprematist basic element, the seed of all the others suprematist forms. His own movement of Suprematism enabled him to construct images that had no reference at all to reality. Great solid diagonals of color in Dynamic Suprematism floated free, their severe sides denied any connection with the real world, where there is no straight lines.

Suppose, for example, that Kasimir Malevich knew about an evolutionary art environment, and decided to try it. How does this kind of system work? Most evolutionary art system tend to resemble each other closely. They all generate new forms or images from scratch (random initial populations). They rely completely upon a human evaluator to set fitnesses for each member of the population – normally based on aesthetic appeal. Population sizes are usually very small (often less than ten individuals), to allow them all to be quickly judged every generation. User interfaces are often similar, with members of the current population shown on the screen in the form of a grid, allowing the user to rank them, or assign fitness scores by clicking on them with a mouse [9].

Let it be an evolutionary art environment such that it generates frames of geometric primitive (line, rectangle, circle, arc) compositions to be evaluated by the user. Suppose that Malevich decided to generate a composition like the Black Square. A possible description for that composition could be: a square frame, with white background, with a black square with size of about 80% of the frame, in the center. Let this composition be called Malevich’. Malevich would have good chances of obtaining a result as presented in figure below, after having selected values (rectangle, white, 1, black, solid) for parameters as (format, background, object number, object color, texture). In this particular environment, it does not exist the “square primitive”, only rectangle. So, the rectangle will have to “evolve” to a square.

Fig. 1 - Resulting frames from an evolutionary art system, instantiated to generate one black rectangle on a white background.

The set of possibilities resulting from this parameter choice in that evolutionary system contains all the black and solid rectangles on white square frames, and this set certainly contains the Malevich’ composition. One could reply that it is possible to simply write a program that attends these constraints. It is really simple, but this is not the intention. Here, the intention is to reward the system whenever it presents a composition closer to Malevich’s Black Square, or any other desired. In case of figure 1, the highest note would be given to the right frame. How many interactions would be necessary to obtain Malevich’ composition? After fifty interactions, the obtained results are presented in figure 2: the process may be boring.

Fig. 2 - Resulting frames from the same evolutionary art system, after fifty interactions.

It is possible, still maintaining the initial goal, to program an evaluation mechanism in the system, or an automatic critic, that rewards the generated figures by attributing a score to a composition, according to its resemblance with Malevich’s Black Square. In this way, it is possible to realize a simulation. In one of the simulations, an evaluation of 9.74 (in 10) was obtained in 18 iterations. But to reach the evaluation of 9.99, 1789 iterations were necessary in one simulation, less than 1000 iterations in another; more than 6000 iterations in still another and surprisingly, in another one, the score 9.9 was obtained with 175 iterations.

The evolutionary art system, a population-based search device, was evolving pictures. A chromossome, described by the selected values (Square, White, 1, Black, Solid) was associated with each picture, as well as others values were associated with the object’s dimensions (in this case, only one square object). The initial pictures, randomically generated, might be closer or farther from the objective picture, the Malevich’ composition. Mutation and crossover operators were applied to the best evaluated pictures, and the results were better or worst. The computational environment did not favour specifically this kind of composition.

Despite of the fact that the evaluation function considered the centralisation of the figure or, in other words, there was a selective pressure [10] to centralise the black square object, there was not any specific information in the chromosome specifically addressing the centralisation, or similarity. In other words, the result was reached without the object’s chromosome have in itself a basic information as “centralised”. The evaluation for the selection was based on the appearance (the phenotype) of the object, no information concerning similarity existed in the chromosome (the genotype), and the result emerged, approaching the objective, simply because of the selective pressure. Even without being part of the genetic code, phenotype characteristics were transmitted to the breed. But, because they are not part of the genetic code, these characteristics can easily disappear.

Fig. 3 Frames generated with the same kind of programming rules.

But, if so hard was to obtain the Malevich’ composition in that system, the same kind of programming rules immediately generated the pictures in Figure 3. Much more complex compositions were “spontaneously” generated. Probably, Malevich would enjoy to explore this kind of composition, it has some similarity with Malevich’s paintings, like the one presented in figure 4. This was the kind of figures for which the system was designed to generate. So, can the computers be creative? The generative potential of a system is not always obvious. The computer can realise tasks that were not explicitly ordered and often fails in realising tasks that were supposed to have been well specified. All the details in the frames of figure 3 were not specified, while a lot of information was supplied for generating Malevich’ (the Black Square on White Background) composition.

3. Creativity and Machines

How compelling are the arguments against machine creativity? The claim that computers cannot be creative turns out to be a cluster of related claims. A common version is that they cannot be creative because they merely follow instructions. But sometimes people are instructed to be creative. Pope Julius II instructed Michaelangelo to be creative when he painted the Sistine Chapel ceiling [11]. So it is possible both to be creative and to be following instructions. The reply to this will probably be that Julius only gave Michaelangelo very general instructions, and left the rest to him, whereas every single thing that a computer does is something that it was told to do. But in fact we do not instruct computers in every action that they perform. This would require us to give them millions of instructions per second. The reply may now be that everything that they do follows from instructions that we give them. But what does this mean? If it means that the machine's performance literally follows from its instructions then it is false, for if we wrote all the instructions on a piece of paper, nothing at all would happen. Presumably it means that computers are designed to respond in a predictable way to their instructions. But even this isn't clear. Does it mean that computers are predictable, in the sense that we can predict their output given their input plus an exhaustive account of their innards? Or does it mean that we have designed the innards, so that the creativity is really ours? There is a general perception that computers cannot be creative.

Fig. 4 Suprematist Painting by Kasimir Malevich, 1915-16. Oil on canvas, 49 x 44 cm (19 1/4 x 17 3/8 in); Wilhelm Hacke Museum, Ludwigshafen

Lady Lovelace said that computers "have no pretensions to originate anything". This, too, has taken root in our culture, so that we tend so that we tend to believe that human beings can be described as being machines, to believe that people are machines, but also that machines cannot do something which is characteristically human. There are in fact two intuitions here. If machines cannot be creative, then (a) they cannot be intelligent, and (b) people (who can be creative) cannot be machines. If machines cannot be creative, then they cannot have "minds of their own", in the sense of being able to generate their own ideas, and it is difficult to see how a system that cannot generate its own ideas can be intelligent. This would be the end of Artificial Intelligence researcher’s aspirations to develop intelligent machines.

But, if creativity is not a computational process, it might still be possible to simulate it computationally [11]. According to Margaret Boden [12], there are three main types of creativity involving different ways of generating the novel ideas. The first, the combinational creativity, involves novel (improbable) combinations of familiar ideas. The second and third types are closely linked, and more similar to each other than either to the first. They are exploratory and transformational creativity. The distinction between an exploration and a transformation is to some extent a matter of a judgement, but the more well defined the space, the clearer this distinction can be. Evolutionary computation is a population-based search engine since it works all the time with a bunch of things, and is certainly a very strong exploration tool. The mutation operator also contributes to exploration. Moreover, because of the crossover, is also combinational and this puts it very close of a transformational tool. So, evolutionary algorithms seems to fit very well to simulate creativity.

Conclusion

Evolutionary art systems supply an environment for creating and exploring novelty, often complex novelty, without requiring human understanding of the specific process involved. Combining evolution and learning can allow a system to “learn” about human aesthetics from the user. Certainly, to teach a system to create visual compositions like Malevich, for example, does not make the system “a Malevich”. But, in the same way, to teach a human to create visual compositions like Malevich does not make him “a Malevich”. But both of them can surprise us with unexpected results.

References

[1] Boden, M. “Creativity and artificial intelligence”, Elsevier Science: Artificial Intelligence 103 pp. 347 – 356, 1998.

[2] Turing, A. M. “Computing Machinery and Intelligence”, Mind, Vol. 59, pp. 433 – 460, 1950.

[3] Todd, P. M. & Werner, G. M. “Frankensteinian Methods for Evolutionary Music Composition” em Griffith, N. & Todd, P. M. (eds) Musical Networks: Parallel Distributed Perception and Performance, Cambridge: The MIT Press, 1999.

[4] Sims, K. “Interactive Evolution of Equations for Procedural Models”, The Visual Computer, 9: 466 – 476, 1993.

[5] Moroni, A., Von Zuben, F.J. & Manzolli, J., “ArTbitration: Human-Machine Interaction in Artistic Domains”, Leonardo, 35(2), pp. 185 – 188, 2002b.

[6] Baluja, S., Pomerleau, D. & Jochem, T. “Towards automated artificial evolution for computer generated images”, Connection Science, 6(2 – 3), 325 – 354, 1994.

[7] Gibson, P. M. & Byrne, J. A. “NEUROGEN, musical composition using genetic algorithms and cooperating neural networks”, London: Proceedings of the IEE Second International Conference on Artificial Neural Networks, pp. 309 – 313, 1991.

[8] Malevich, K. http://www.ibiblio.org/wm/paint/auth/malevich/sup/malevich.black-square.jpg, 1913.

[9] Bentley, P. & Corne, D. (eds.) Creative Evolutionary Systems. San Francisco: Morgan Kaufmann, 2002.

[10] Dawkins, R. The Blind Watchmaker. London: Penguin Books, 1991.

[11] Dartnall, T. (ed.) Artificial Intelligence and Creativity: an Interdisciplinary Approach. Dordrecht: Kluwer, 1994.

[12] Boden, M. “Creativity and artificial intelligence”, Elsevier Science: Artificial Intelligence 103 pp. 347 – 356, 1998.

Acknowledgement

I am very grateful to Dr. Fernando José Von Zuben for helping me reviewing this paper. I would like to thank very much Daniele Gugelmo Dias for her strong support in the presentation of this paper.