Understanding Music with AI: Perspectives on Music Cognition

Rating:
Summary: Preliminary to subsequent research in machine music
Review: Art and music used to be thought of as two fields of human endeavor that could never be realized by artificial intelligence. That belief is still held by many, but in the past few decades painstaking and dedicated research in artificial intelligence has shown beyond doubt that not only can non-human machines compose music that is beautiful to listen to, but one can use these machines essentially as tutors, giving keen insight into musical composition and music theory. The musical expertise non-human machines allows a deeper and richer appreciation of music, and the music they produce will continue to stir the senses and interrupt, or perhaps dominate, the normal course of cognition.

Via a collection of research articles, this book gives a splendid representation of what was done in using the field of artificial intelligence to understand music theory and composition up until the year 1992. The last twelve years of course, thanks mostly to faster and more powerful hardware, has seen considerable advances in musical artificial intelligence. The quality of music composed by the machines is astounding, and considering that hardware is continuing to get more powerful (and cheaper), it will be interesting to see what the musical abilities of the machines will be a decade from now.

The book essentially defines itself as an overview of 'cognitive musicology', which as Otto E. Laske asserts, is a field that began in the 1970s, and has as its goal the understanding of both musical thought and 'musicological' thought, and their links to 'musical action.' It has its origins in many different fields, such as cognitive psychology, neuroscience, artificial intelligence, and semiotics, and attempts to model musical knowledge, but does so in a way that does not separate knowledge from action. Laske wants to move away from the Cartesian paradigm, believing that it is inadequate for music research. He also believes, interestingly, that there is a 'musical intelligence' that is distinct from various other types of "intelligences" that can exist in humans. Thus cognitive musicology should be viewed as a field that studies this musical cognitive system, and this study can be done independently of the research into other forms of intelligences, such as linguistic or mathematical.
Laske breaks up the field of cognitive musicology into: 'local knowledge', which is knowledge about the tools and materials needed; 'competence', which is knowledge about the domain; and 'performance,' which is knowledge of how to perform under real-time constraints. The integration of work in cognitive musicology with computing machines is essential according to Laske, for this will allow the view of music and musicology as essentially knowledge engineering. Artificial intelligence is and essential part of cognitive musicology he argues, since it introduces a task-oriented perspective on music, which had not been done in music theory at that time.

The article by Peter Kugel follows the one by Laske, wherein Kugel argues that the strict computational framework of Laske may be inadequate for some forms of musical thought. To make his case on the limitations of computation, he introduces what he calls an 'announcement condition.' This is a method by which one can tell with certainty whether a procedure has finished doing its job. This motivates the idea of a 'limiting computation', which is one that allows a solution to a problem that a "regular" computation can't. Musical thinking, Kugel asserts, requires limiting computations, and he discusses various methods for finding examples of musical thinking that require more than regular computations. Interestingly, Kugel uses Cantor diagonal arguments to find, or at least indicate how to find, examples of new styles of composition. There are problems he says that can be found by "technique", but others require "insight", and the difference between these does involve the level of knowledge of the problem solver. One can turn some problems requiring insight into ones that do not, but there are some problems, such as those involving musical creation, that cannot be. No explicit examples are given however.

Many other very interesting articles follow, all discussing various aspects of how to model musical activity, connections with artificial intelligence, automated musical composition, etc. One particularly interesting article is the one by Kemal Ebcioglu on designing an expert system for harmonizing chorales in the style of Bach using first-order predicate logic. Written first in PROLOG on a VAX 11 architecture (which shows the age of the article), Ebcioglu describes how a language called BSL (for Backtracking Specification Language) was designed for the purposes at hand. The language was constructed so as to permit the coding of universal and existential quantifiers, be efficient for producing high-quality music, and yet be tractable and easy to use. An illustration of the syntax of the language is given using the 8-queens problem and an informal description is given of the semantics of the language. The search technique of backtracking plays, as the name of the language implies, a central role, but it is implemented in a manner that makes it less inefficient than the usual backtracking techniques that are implemented in LISP and PROLOG. The author then describes the CHORAL system, which allows the representation of Bach chorales and their harmonization.

Rating:
Summary: Pioneering investigations of musical behavior
Review: Understanding Music with AI is an introduction to Cognitive Musicology, the term understood as a science of mental representations of music. The book introduces formal models for otherwise taken-for-granted musical activities such as composing, analyzing music, listening, performing music, etc. In contrast to present-day work with neural nets, theories introduced in the book test their models at the level of cognition, rather than perception or brain processes. The distinction between listening and perception, e.g., is thought to be fundamental, in that listening is based on meaning-making, and comprises perception as a mere subprocess. There is also an attempt to do justice to creativity, especially in music composition. As a consequence, criteria of theoretical adequacy are different in AI- and network-based music research. For a critique of the premises of AI work, see Marc Leman, "Adequacy criteria for models of musical cognition," in J.N. Tabor (Editor), Otto Laske: Navigating new musical horizons, The Greenwood Press, 1999, pp. 93-120. Readers not interested in these academic matters will enjoy reading the book for the intricacy of the systems displayed therein, and the pioneering spirit of the contributors. Otto Laske, Co-Editor (1992).