Subscribe to our mailing list to receive updates about forthcoming issues. JCMS will not disclose your details to any third party.
Volume 1, Issue 2, March 2017
CSMC16 conference review, computer simulation of musical creativity, evaluation of computational creativity, human-computer co-creativity
The 1st Conference on Computer Simulation of Musical Creativity (CSMC16) was held June 17–19, 2016 at the University of Huddersfield. Several themes emerged in the conference discussions addressing some of the fundamental questions of what “creativity” is, or could be, as well as issues regarding methodologies for evaluating the potential “creativity” of computational systems. For this conference review I invoke Wiggins’s (2006) formalisation of creative systems, understood as searches in conceptual spaces, and I use this as a working understanding of creativity in order to suggest some questions related to the papers and discussions; these questions may provide fuel for themes in future conferences in this cross-disciplinary field.
The 1st Conference on Computer Simulation of Musical Creativity (CSMC16) was held June 17–19, 2016 at the University of Huddersfield. CSMC16 attracted some 40 delegates and provided a platform for a lively exchange of ideas from researchers who represented a wide range of specialisations from the fields of computer science, musicology and music composition. This multi-disciplinary forum provoked exchanges of ideas that, judging from several participant comments, enlarged the delegates’ perspectives both within their own fields as well as in the cross-disciplinary realm. Some themes emerged in the conference, especially regarding problems of understanding what “creativity” is, or could be, as well as issues regarding methodologies for evaluation of the “creativity” of computational systems. It seemed that the methodological problems were rooted in a lack of clarity about what constitutes “creativity”. There was also questioning of whether computational creativity should, or even could, be compared with human creativity. Perhaps this kind of comparison is a category mistake?
For this context I will evoke the idea of “creativity” in the sense formalised by Geraint Wiggins, as explorations in, and transformations of, “conceptual spaces”. Wiggins’s formalisation, based on the ideas of Margaret Boden (2004), described two levels of creativity within his formal system as: 1. “exploratory creativity” and 2. “transformational creativity” (Wiggins, 2006, pp. 453–454). But how should such explorations or transformations be represented in order to achieve computationally viable implementations? Peter Gärdenfors proposed geometric representation of “conceptual spaces” as complementary to the symbolic and associationist (including connectionist) methodologies for modelling representation (2000, p. 1). If these three approaches “attack cognitive problems on different levels”, as argued by Gäardanfors (Gärdenfors, 2000, p. 1), then the question arises: could a given implementation, which does not apply all three approaches,(1) be considered a partially creative system? This possibility was implicit in some of the CSMC16 papers and I will return to this point in Section 2, as it may be a useful consideration for asking questions that might contribute to methodological development of computational creativity evaluation.
It may well be that there is no unitary process of creativity,(2) in other words no “one coherently conceptualizable phenomenon for which there could be a single, conceptually coherent [approach or model]”. Instead different parts of what we call “creativity” could “share what Wittgenstein called ‘family resemblances’, that is, some overlapping similarities”. Thus “creativity” could be understood “as a family name for a series of capacities that have some overlapping similarities as well as some significant differences” (Cunningham, 2000, p. 67).(3) In other words, the three methodologies – conceptual spaces, symbolic computation and associationism – may all describe different parts of a creativity concept family which can be relevant for computational systems. It is significant to note that the creativity concept families for humans or other living organisms may contain parts that do not correspond to those involved in computational systems, and this was implicit in some of the arguments for a separate category of computational creativity, during the second day of the conference, to which I will return in Section 3.
The first session of CSMC16, entitled “Machine-Learning Generative Systems”, addressed the use of “long short-term memory” and “recurrent neural networks” for “deep learning” with a large corpus of Irish folk songs presented by Colombo et al. (2016) and Sturm et al. (2016), or Jazz standards presented by Choi et al. (2016), and with subsequent applications of the systems as melody generators or “automatic composition systems”. Artificial neural networks are, of course, central to much connectionist-based work, but Sturm et al. “make no claims that [they] are modelling music creativity” (Sturm et al., 2016, p. 14). However, the training and subsequent generative output of the artificial neural networks may well constitute a cognitively necessary (but not sufficient) part of a creativity concept family of computational creativity (as discussed in Section 1).
Colombo et al. suggest that their neural network melody output “could be evaluated by humans in a Turing test setting” (Colombo et al., 2016, p. 10). This implies confidence in a probabilistic belief-structure: that statistical knowledge of a melodic repertoire will be sufficient to generate the expression of a new melody, within the same genre, which can be considered a “creative” output in human terms. I would add a cautionary note that, if the melodies are evaluated on the basis of the same probabilistic belief-structure (using only reference to the training repertoire), then there is the possibility that evaluation becomes a circular process, entangled with the generative approach. In other words, the process can only operate in Wiggins’s “exploratory creative system” within a concept space that is limited by the training repertoire; if evaluation of the results is also limited by the same concept space then creation and evaluation appear to be entangled.
Anecdotally, I have heard a number of creative musicians express the desire to perform their own “voice” within a genre (such as modern jazz), and this seems to imply that they seek to add something external to the conceptual space of the established repertoire or musical idiom: something unique to their own practice, and therefore something not previously found in that repertoire or musical idiom. This arguably means that a search within an idiom-limited concept space may not be sufficient for human creativity; instead it is a partial creativity, and we might see it as one part of a larger creativity concept family.
The second paper session, entitled “Computers as Creative Tools”, was focused around computational music analysis of common-practice repertoire to investigate musical style (White, 2016) and sound analysis of vocal techniques that was used as a pedagogical tool (Bonin, 2016). These presentations provided analytically applied technologies, giving contrasting perspectives to the more generative orientations of the previous discussions. If we accept analysis as a creative, or re-creative, process,(4) then computational tools for analysis might enter a co-creative relation with the human analyst; I will discuss co-creativity further in Section 4.
On the second day of CSMC16, in paper session 3, entitled “Theoretical Perspectives on Creative Music Systems”, the evaluation of creativity in computer systems came to the forefront again. The presentations included arguments for machine evaluation of systems presented by Loughran & O’Neill (2016) and a taxonomy of a “General Creativity” wherein machine-generated music would be considered a superset of human-made music presented by Velardo.(5) Loughran & O’Neill proposed to move away from a human standard for evaluating computational creativity and instead move towards a computational standard, stating that “[a] more interesting system is one that can discover itself what it believes to be beautiful [music] – not what it is told is beautiful” (Loughran & O’Neill, 2016, p. 12). In a similar vein, Velardo proposed a taxonomy of creativity, where “musica humana” would be a subset of “musica mecanica”, in the sense that computationally generated music would include more possibilities than human-made music. Both authors proposed that computationally generated music could be valid for a computational listener (or Artificial Intelligence), even if the music might be rejected by human evaluation.
These presentations sparked some controversy and hence lively discussions in the friendly atmosphere of the conference. The anthropocentric approach to evaluating computational systems was challenged, and this seems to be an area which is in need of further work. It was argued that human and computational evaluations might be complementary and could coexist, but also that music is an essentially human enterprise and so must be evaluated by human criteria. However, there was also a resistance to defining music, and it was argued that the human-centric idea of music may at this point be too defining, or too limiting. Perhaps our understanding of what music could be should be expanded given the context of current creative systems?
As mentioned in Section 1, the creativity concept families for human activity may contain parts that do not correspond to those which could potentially be involved in computational systems. This would seem possible, at least if we argue that the human mind is not a purely computational system. Whether the human mind can be fully described as a computational system is an open question. But, if the comparison between human creativity and computational creativity is a category mistake, then the arguments for separate approaches to evaluating these “creativities” (or classes of creativity) would seem to be supported. However, the implications from Loughran & O’Neill that a separate computation-based aesthetics is necessary, or from Velardo that such a “musica mecanica” aesthetics would be a superset of human aesthetics, do not seem to follow. It may be that there is a need to expand the field of aesthetics, in order to accommodate possibilities made available by computational creativity, but as a philosophy of art, the field of aesthetics would not seem to be dependent on the capabilities or characteristics of any particular agent, human or otherwise. I will not attempt a thorough investigation of these issues here, but a future CSMC conference theme might fruitfully address some of these issues in the cross-disciplinary area encompassing aesthetics, philosophy of mind and philosophy of computational creativity.
The morning paper session of the third day of CSMC16 had the theme “Computers as Composers” and provided three different approaches. The first described the ideas behind the implementation of an “improvising” machine accompanist, and the analytical evaluation of the music that resulted from testing the accompanist with human improvisers (Mogensen, 2016). The second presented a preliminary rule-based software implementation for generating species counterpoint (Ren, 2016). The third presented concepts for developing computational tools to assist in sound design (Cherny et al., 2016). These could all be understood as human-computer collaborative systems for generating solutions to predefined musical problem-areas: improvisational accompaniment, species counterpoint and sound design. As such, these projects might enter Wiggins’s formal framework as examples of “task-divided co-creativity” in the terms of the extension of the framework proposed by Kantosalo and Toivonen (2016). An implication of the session papers was that partially creative computational agents could complement creative problem-solving (or performance) by humans, and thereby be co-creative. One could speculate that partially creative agents in a co-creative system might contribute sufficiently for the system as a whole to have some kind of emergent co-creativity (greater than the capacities its individual contributors). But would a necessary condition for this emergent co-creativity be that the agents provide a sufficient complement of parts for a complete creative concept family?
From a musicological perspective I argue that we can draw some parallels between the computational creativity discussed here and 20th century “integral serialism”.(6) One common thread might be the goals that arguably appear in both of these directions: to create codified generative systems for music that are potentially interactive with human musical sensibilities. The motivations of the approaches to music are comparable: for “composers who adopted total [integral] serialism in the early [Nineteen-]Fifties such as [Luigi] Nono, [Bruno] Madera, and [Franco] Donatoni,… the main reason for the use of predetermined principles was to obliterate memory” of the musical culture that was associated with the cultures leading to the two World Wars (Brindle, 1993, p. 23). Integral serialism was “[t]he avant-garde composer’s ‘fundamental revision of every element of the musical language’” (Brindle, 1967, p. 161). This resonates with the current arguments for a distinct approach to evaluation of computationally created music (discussed in Section 3), based on the idea that such music is somehow a distinct class of music since it has been generated from a distinctly computational class of creativity. The systems of “integral serialism” were arguably co-creative with the composers who implemented these systems on paper as music scores. Given these parallels, what might be the intertextual(7) relations between current work in computational musical creativity and “integral serialism”, or indeed other systematic composition or performance tools?
A “Round Table” entitled “Creative Music Systems: Current Capacities and Future Prospects” opened the floor to a general discussion. A number of current systems were discussed, but perhaps more significantly this session resulted in some degree of self-reflection by the conference delegates as a group. The cross-disciplinary demography of the delegates appeared to support intellectual synergies in the discussions, the results of bringing together researchers from diverse fields who had at least three common interests fundamental to their work: music, creativity and computation. The potential values of exchanges between generalised and specific computational solutions were questioned, and the problems of evaluating “creativity” surfaced again. The potentially hybrid nature of human-machine activity was brought up and the pervasiveness of the cyborg-aspect of contemporary human life was discussed. Given the context I have outlined in this review, I would suggest a related question for future conference reflection: what are the necessary and sufficient requirements for hybrid human-machine activity to be co-creative and how does this differ from, or coincide with, the requirements for other potential concept families of creativity?
The final keynote was given by Wiggins, who offered some “reflections on computational creativity in music”. In his very stimulating talk, Wiggins presented recent collaborative research and proposed that “creativity is prediction in the absence of input” in a potential answer to what had emerged as a theme question for the conference. The idea of “prediction” came from approaching creativity as a function of memory, apparently somewhat similar to generative use of deep learning, but potentially based on higher-order statistical processes. It appeared to be a very promising research direction, although it seemed to remain within the “exploratory creativity” of Wiggins’s (2006) formal framework. I wonder whether “prediction” can become the basis for modelling “conceptual leaps” in human thought? Could “prediction in the absence of input” have resulted in the development of 20th century “integral serialism”? In other words, could statistically based decision-making enter the realm of the formal framework’s “transformational creativity”? Or, alternatively, are “exploratory creativity” and “transformational creativity” perhaps two distinct parts of a creativity concept family?
This brief review of CSMC16 has been biased by my own interests in the many open questions around creativity, computational creativity and its evaluation, and so I have not mentioned all the conference papers or workshops: my apologies to those I have omitted.(8) The questioning around creativity and evaluation of creativity seems fundamental to our understanding of our own humanity as well as our present and potential future relations with computational partners or tools. To give a working understanding of creativity for this review, I have invoked the formal framework for creativity proposed by Wiggins, but this model has some limitations: it is intended to allow “detailed comparison, and hence better understanding, of systems which exhibit behaviour which would be called ‘creative’ in humans” (Wiggins, 2006, p. 449). Thus, it entangles human and computational creativity and does not provide a clear basis for understanding the possibility of multiple “creativities” applicable distinctly to computational systems or humans. Also, if there is a computational creativity that can be considered distinct from human creativity, then what are the implications of this potential distinction for co-creativity between humans and computers? These questions could be fruitfully addressed during papers and panel discussions in future conferences, to move further towards a broader working understanding of creativity in the context of computation.
The delegates of CSMC16 ranged from PhD students and early-career researchers to more established researchers, and the atmosphere of the conference gave the impression of a well-tempered group of peers who challenged each other in their searches for knowledge. The cross-pollination of ideas from the fields of computer science, music composition, musicology, philosophy, as well as input from the software industry, made the conference an enriching and stimulating experience. Future editions of this conference promise to be catalysts for further cross-disciplinary exchanges and developments.