Musical Metacreation

点击翻译按钮获取中文摘要

From flying planes to playing Jeopardy, Artificial Intelligence has been tremendously successful at rational problem solving at a human competitive level. Computational creativity is the fast growing field devoted to automatizing creative tasks, that is endow machine with creative behavior, as a complement to rational problem solving. Creative tasks differ from typical rational problem solving in the sense that the notion of optimality is ill defined. While the search spaces are well defined, there is no such thing as an optimal oil painting, portrait drawing, music composition, interpretation of a composition, level for a video game, narrative, poetry, joke, etc. Musical Metacreation is the sub-field of computational creativity applied to computer music. In the last 5 years, our group – the MAMAS laboratory – has been successful at developing models and systems to generate melodies, harmonic progression, rhythmic patterns, soundscape compositions, and full-fledged electronic dance music pieces, to name a few. All of these applications rest on a collection of new and hybrid algorithms that integrate human composer expertise, computer music and music information retrieval techniques, evolutionary computation, and corpus-based machines learning. Most of these systems have been shown (in statistically significant ways) to be human competitive, that is to be indiscernible from human creations produced under similar constraints. The proposed research project builds on these promising premises, with the following objectives which are direct answers to the current challenges in the field: 1. Develop algorithms for Musical Metacreation that are capable of learning multiple levels of musical structure (pitch and onsets, phrases, parts, pieces, …). 2. Research the high level variables and cross-modal mapping strategies that shall be used in order to steer the musical generation in interactive contexts (computer-assisted creativity in digital audio workstations, adaptive online generation in Video games, …). 3. Further evaluation methodologies for computational creativity. The scientific approaches deployed to tackle these three objectives will respectively be: 1. Developing machine learning algorithms to capture the overall horizontal (temporal) and vertical (harmonic and timbral) dependencies in music. To this effect, we will adapt, apply and compare various machine learning paradigms such as: evolutionary computation (GP, Cartesian Programming) and hierarchical, multi-level hidden Markov Models, recurrent neural networks, and Deep Learning algorithms which can - in principle - learn several levels of abstraction in an unsupervised fashion. 2. Exploring various ways to parametrize these models and develop high-level cross-modal mappings. For example, we will build on our award-winning previous work on video game level generation and use the tension curve of the game level as well as the situation of the player to control the music generation. 3. In order for these systems to be validated and later adopted by the industry, one need to show that they are human competitive. We have, and will continue to contribute to this challenging need by developing appropriate research instruments resting on quantitative empirical experiments that go beyond the musical Turing test. As non-linear medium develops, the number of assets needed is growing exponentially. Beside its contribution to artificial intelligence, machine learning, computational creativity, computer music, and music information retrieval, computer assisted or completely procedural music generation is now required in a variety of software products (interactive systems, video games, ...) that are central to the digital economy and Canadian industry in these sectors.