Creativity is a defining property of humanity. All humans can create, but few other species reliably exhibit creative behaviour. Human creativity is highly valued, and we claim to educate our children so as to foster it. Creative behaviour, however, is not only a matter of genius or rare talent: it is exhibited on many levels, from great works of art in opera, painting and sculpture, to everyday, small-scale generation in forming the next sentence in a conversation, or working out how to open a sealed package in the absence of scissors. Creativity is fundamental not only in artistic practice, but also in science and engineering: without creativity, hypotheses cannot be formed, experiments and software cannot be designed, bridges cannot be built. In short, without creativity, there is no progress.
It is therefore puzzling that, until recently, there was almost no attempt to study creativity scientifically, and none at all in the context of Artificial Intelligence (AI). In the 20th century, philosophers and psychologists such as Wallas (1926), Guilford (1967), Koestler (1976), Getzels and Csikszentmihalyi (1976) and Boden (2004) proposed theories and mechanisms, and some computational attempts were made to solve restricted creative problems (e.g., production of particular restricted kinds of music: Ebcioğlu, 1988; Cope, 1987; Pachet and Roy, 1998; Wiggins et al., 1999; styles of architecture: Mitchell, 1992; or kinds of mathematical proof: Bundy et al., 2005). However, these attempts were almost always cast as problem-solving activity in their own domain, often in a human-coded, rule-based style, lacking an underlying, explanatory theory to account for the creative behaviour associated with the equivalent human activity. Only in the mid-to-late 1990s was creativity itself studied directly, in the AI context (Boden, 1999; Cardoso et al., 2010). Even once this step-change in approach was achieved, the tendency to focus on restricted problems (imported from standard engineering methodology) persisted, with the consequence that it is difficult to argue that any of the systems thus produced are actually exhibiting creativity in a holistic way, or that they elucidate the human equivalent. A literature survey reveals fragmentation and the lack of a generally agreed conceptual framework to bind the findings of research in several disciplines into a coherent body of knowledge.
Creative behaviour provides the flexibility and adaptability needed for any intelligent organism, human or artificial, to act meaningfully in a dynamic environment. For example, successful generation of creative solutions is applicable in autonomous remote search-and-rescue systems and remote exploration systems, where unanticipated difficulties can arise. But many other application domains (consumer robotics, production, construction, software engineering) will also benefit from the increased robustness of creative systems over computer systems that follow fixed operation sequences. Furthermore, at higher cognitive levels (i.e. culturally, and scientifically), artificial systems will need creative capacities in order to act on a par with humans. To be valuable to humans in a given context (e.g., in live music performance, or in scientific debate), such systems must make novel contributions that are meaningful in that context. The related capacity to judge novel contributions by their aesthetics or function could play a major role in computer tutoring systems.
An underpinning faculty for any creative system is the capacity to anticipate; indeed, creativity itself has been characterised as the ability to anticipate the unexpected (Koestler, 1976; Boden, 1990; Shanahan, 2010), and then to identify value in the anticipated artefact (Boden, 1998; Shanahan, 2010). The evolutionary imperative for anticipation in biological systems is self-evident: an organism capable of anticipation is better able to survive in the world than one which is not. For example, given the capacity to anticipate, an organism can orientate itself optimally by predicting what is likely to appear in a given context (Zajonc, 1968; Huron, 2006), react more quickly where necessary (Schultz et al., 1997), and comprehend partial, noisy or ambiguous stimuli by anticipating unclear or missing components (Summerfield and Egner, 2009).
In order to anticipate, organisms must know what to expect. Therefore, they must learn, and they must learn implicitly, by mere exposure (that is, without being told: Zajonc, 1968); otherwise, they will not be able to associate antecedent with consequent in a changing world. This is the kind of learning by which children acquire native language (Saffran et al., 1997); indeed, humans are so sensitive to structure in perceived stimuli that we are even able to learn the theoretically-inaudible structure in bursts of white noise (Agus et al., 2010). Associations between real-world objects and words underpin semantics, while associations between words in sequence underpin syntactic structure (Ge and Mooney, 2005). The whole leads to a system in which positive feedback allows the evolution of novel linguistic forms (Kirby et al., 2008)—a kind of social creativity.
However, learning through experience is not enough: an organism cannot usefully learn to avoid fatal consequences by direct experience. This practical necessity creates evolutionary pressure to develop emotional responses—preferences—related to the realisation or denial of expectations, so that fatal consequences can be avoided without direct experience (Huron, 2006). These preferences are likely to contribute to aesthetic experiences related to creative activity (Pearce andWiggins, 2012), and in turn to preferences between, and value attributed to (Boden, 1998), created artefacts.
Given the generally accepted implication of intentionality in the common usage of “creativity”—that creators are aiming to create, not doing so by accident, which would be serendipity, rather than creativity—knowledge, implicit or otherwise, of the domain in which creativity is taking place is paramount; one cannot intend to create something unless one can imagine what it might be (though it is important to emphasise that such knowledge can still be implicit, and often is so in human creativity). For this reason, successful theories of creativity include a notion equivalent to that of Boden’s (1990) conceptual space: a closed, but unexplored, set of concepts which are potentially available to an entity creating in a given domain. Given this simple philosophical tool, creativity mechanisms can be straightforwardly characterised in terms which are suitable for implementation on computers (Wiggins, 2006a), measures can be proposed, helping to define what may be “more” or “less” creative (Ritchie,
2001), and interesting effects at the edge of the conceptual space can be formalised and detected, simulating, for example, what happens when a creator changes the membership of its personal conceptual space (Wiggins, 2006a). This conceptual space may be thought of mathematically as a complex prior distribution, learned by exposure, whose probabilities represent co-occurrence and/or correlation constraints on the features of the entities in the conceptual space; thus, it corresponds with the result of the learning process for which we argue above.
Wiggins’ (2006b) formulation is that creativity at the non-conscious, basic level can be viewed as non-conscious enumeration of possibilities, with selection of high-quality choices causing conscious awareness. Given the well-established massively-parallel nature of the brain, it seems highly unlikely that this enumeration would happen in series, and, indeed, current theories of consciousness, most notably that of Shanahan (2010) support the idea of perpetual generation of material, in particular linguistic utterances, at a non-conscious level, with mechanisms for selecting some according to relevance or other criteria, which are then made conscious by well-defined mechanisms. In the proposed project, we explore the relationship between creativity and a radical new theory of conscious attention, which is related to information theory, learned knowledge, and, therefore, memory. The current proposal, therefore, is placed at the centre of cognitive science, proposing a uniform, testable, explanatory theory for the operation of mind, from which the faculty of creativity naturally emerges.
This must be distinguished from the theoretical construct of the same name due to Peter Gärdenfors (2000); the two are not identical, though they can be unified (Forth et al., 2010). Gärdenfors’ ideas are relevant elsewhere in this proposal.
Ultimately, any creative system, human-like or otherwise, is likely to include the following components, or proxies for them (Koestler, 1976; Boden, 1990; Wiggins, 2006a):
1. a mechanism for perceiving the creative domain;
2. a mechanism for learning about the creative domain;
3. a mechanism for generating examples in the creative domain from a learned model; and
4. a mechanism for evaluating generated examples according to both novelty and value.
In this project, we focus on the last two (while bearing in mind the need for the first two), and in particular on appropriate knowledge representations and associated generation methods. These are placed in context of a cognitive architecture specifically designed to admit creativity as a normal human cognitive function. We will evaluate these systems scientifically and report on the methods and technology developed. The overall scientific approach of the project will require radical new research methods both at the technological and the methodological level, which constitute our first and most general objective.
We aim, then, to further exploit the potential of computational resources for society, by enhancing them with creative capacity. This aim adheres to a long-term vision of increased general utility of computers by using them not only as advanced tools (to be operated explicitly by humans), but also as both collaborative and autonomous entities. Such entities should, on the one hand, interact with and support human beings at high cognitive levels, and, on the other hand, be capable of performing complex tasks in dynamic environments without human intervention.
A necessary condition for rigorous science in this area, including in the proposed project, is the development and documentation of proper scientific approaches appropriate to the computational study of creativity and resulting information technology, especially evaluation. This proposal is itself such an approach, but it will be necessary to design and develop rigorous methods and, in particular, rigorous evaluation techniques for creative systems and their outputs as part of the project’s activity.
Before a computer system can be unequivocally shown to exhibit creativity, it must be able to learn the nature of the domain in which it operates; otherwise, the counter-argument can always be made that a programmer has been creative, instead of the program itself (Jennings, 2010). Learning in general, however, requires appropriate data.
A major impediment to the development of creative systems is that the vast majority of creative acts are located inextricably in a real-world context. To see this, consider the creation of narratives: to create a non-trivial narrative, one must have a model of the world (real or otherwise) in which the story takes place; good comprehension of that world, and all the connotations it encodes, is a crucial part of creative success. For example, a narrative concerning an event as ordinary as a football match involves extensive understanding of ballistics, game rules, team spirit and related cultural mores, physical capabilities of footballers, and so on. The same problem arises for figurative visual art: the limitations of designing scenes are so great that the longest standing researcher in creative systems for art, Harold Cohen (1988), recently abandoned scene construction for abstract painting, with his computer painter, AARON (McCorduck, 1991). The AI models required for such rich real-world reasoning are not yet available; nor are they likely to be in the near future. Notwithstanding various excellent attempts since the early days of AI (e.g., the Naïve Physics Manifesto: Hayes, 1979), rule-based models of how people think about the world are not adequate for realistic creative reasoning: the CYC project, the largest effort in general reasoning to date, is still unfinished, 25 years on and with an estimated 325 years to go (Guha et al., 1990). In very recent years, the arrival of the Semantic Web (SWeb) has begin to make an impact on this problem; it addresses the problems inherent in research programmes of this nature in two particular ways:
Person-power The amount of information required to describe human knowledge of the world is unimaginably large, and growing. The SWeb potentially leverages the person-power of all Web users—a significantly larger resource than is available to any research centre.
Representation The SWeb has minimal, but adequate, representation standards, which are internationally managed, and logically well-defined. In particular, individual users can create their own ontologies (i.e., knowledge representation languages), and so are able to build what they need, but these new ontologies may be plugged into extant ones, by means of relations between common or similar concepts. Thus, the basic representation is conservatively extensible, and so is unlikely to change in future.
Therefore, working with basic SWeb representation technology (enhanced to admit the specialist kinds of reasoning that form the centre of the project) allows the proposed project access to future knowledge encoded and exposed on the SWeb, while also providing a very neat set of base standards for sharing of data and processes in our consortium.
In any case, from a methodological point of view, to take an absolute position on learning, and to discount top-down guidance or even rule-building would be an unacceptably risky research strategy: failure at any point along what would be a very narrow chain of developments would jeopardise the entire project. Therefore, we mitigate risk by judicious and appropriate study of more deliberately guided methods. This has two effects: first, it allows us (temporarily) to bridge gaps in the research chain from basic learning up to creativity, but also, in cases where learning might not succeed, it provides hypotheses as to the answers that need to be sought: the guided methods become targets for learning research. This presents a rigorous cognitive modeling framework according with that proposed by Wiggins (2007, 2011).
The ultimate aim of our proposal is the production of a computer system that is capable of conceptualising new ideas, in terms of, but not restricted by, its existing knowledge. We will leverage existing symbolic (logic-based) theories of knowledge representation, combining them with non-symbolic theories, as found in, for example, artificial neural networks, and crossing the boundary between the two, via the theory of Conceptual Spaces (Gärdenfors, 2000).
While cognitive architectures are studied extensively in cognitive science and artificial intelligence, there is currently no candidate for a detailed model of creative cognition, nor a theoretical explanation for the existence of creativity as an identifiable cognitive feature. We aim to implement a new model and study its success as a model of human creativity, but also as a creative entity in its own right.
Evaluating creative artefacts and systems is notoriously difficult. The key issue is subjectivity of definition of the terms involved. Statistical methods from psychology can be brought to bear in this area, but a problem remains of terminology and perspective, which are opaque. Currently, the only really satisfactory evaluation method for creativity is the Consensual Assessment Technique (Amabile, 1996). We will design and instantiate new methods of evaluation for this project, founded in behavioural study and in the responses of users of software distributed by our exploitation partner, Chatterbox Analytics.
Appropriate evaluation methods are fundamental to proper scientific study; without them, new information technologies based on these ideas cannot be assessed. Our aim is to make a significant step forward in the evaluation of creative systems, so that the field of computational creativity will be afforded a firmer philosophical and scientific base for its future work.
The ability to behave creatively is arguably the most significant trait that humans (and a small number of other animals) have evolved throughout history. Creativity, defined as the ability to display novel and useful behaviour, is the driving force of progress, both in the arts and in science. The overall aim of this proposal is to assemble the technology and expertise to explore these complex and advanced applications, to develop them further, and to examine and evaluate them from a rigorous scientific perspective.
These outcomes serve a long-term vision of computer systems that can behave in ways comparable with human creativity, both autonomously and interactively, enabling better interaction between human and machine, better autonomous systems in general, and possibly creativity of new kinds, not currently exhibited by humans. We anticipate on-line learning environments which can behave creatively, not only to teach or support creative pursuits, but also to promote creativity in humans. We anticipate immersive gaming and edutainment systems that respond creatively to their users’ actions. We anticipate reasoning systems that can propose new technology that, crucially, had not been imagined by their designers. All of these things become possible with computationally-creative reasoning, once the necessary domain knowledge is made available. Our proposal uses Semantic Web technology to avoid the bottleneck of domain modelling, so that creative reasoning can be ready in advance.