"TSD::AIP1::\"Len@HEART-OF-GOLD\"@atc.bendix.COM" (10/26/88)
Date: Wed, 26 Oct 88 12:00 EDT
From: Len Moskowitz <Len@HEART-OF-GOLD>
Subject: Review of the First Workshop on Artificial Intelligence and Music
To: "3077::IN%\"AIList@ai.ai.mit.edu\""@TSD1
Message-ID: <19881026160039.2.LEN@HEART-OF-GOLD>
The First Workshop on Artificial Intelligence and Music was held on August
24, 1988 during AAAI-88. It brought together more than 40 researchers from the
U.S.A, Canada, Belgium, the U.K., and Israel. The workshop was divided into
five sessions: expert systems; tutoring systems and languages; cognitive models
and knowledge representation; neural networks and parallelism; and a final
session covering perception, philosophy, and the symbiotic relationship between
music and artificial intelligence. Many of the presentations included
audio/visual demonstrations.
The workshop was sponsored by AAAI and organized by Mira Balaban (Ben
Gurion University, Israel), Kemal Ebcioglu (IBM Thomas J. Watson Research
Center), Marc Leman (University of Ghent, Belgium), and Linda Sorisio (IBM Los
Angeles Scientific Center).
From an AI perspective, the workshop spanned a wide range of topics
including planning, machine learning, neural networks, tutoring systems,
knowledge representations, languages, parallelism, pattern recognition,
temporal reasoning, design, and expert systems. From a music perspective, the
presenters focused on analysis of tonal and atonal music, composition, music
education, music perception and cognition, performance, automated
accompaniment, and user interfaces.
This year's AAAI had a noteworthy event. Thanks to the AI and Music
workshop and Harold Cohen's invited talk ("How to Draw Three People In a
Botanical Garden," part of AAAI proper), this was the first time, to my
knowledge, that research carried out within a humanities context received
significant attention.
Many of the attendees expressed a desire for a similar workshop at next
year's IJCAI/AAAI gathering. Anyone interested in organizing or assisting next
year's workshop should contact Dr. Ebcioglu (kemal@ibm.com).
Copies of the workshop's proceedings are available from AAAI (445 Burgess
Drive, Menlo Park, CA 94025, U.S.A.) for $20 (U.S.) plus $2.40 shipping and
handling.
A summary of the sessions follows:
Otto Laske (New England Computer Arts Association) delivered an invited
talk. Laske is perhaps the father of the AI/music synthesis and the field he's
dubbed cognitive musicology. He differentiated between traditional musicology
and the developing discipline of cognitive musicology in that the first is
narrowly centered around the artifacts that musical creates, while the latter
seeks to understand music as one of the processes and structures resulting from
man's 'being in the world.' Viewed in this way, the programs we write are
formal speech accounts of an activity (referents to a human activity) but not
that activity. While programs are not music, they can help us to understand
music. Therefore, if I understand Laske's viewpoint correctly, the goal of the
AI/music synthesis is to investigate, within the music framework, how we become
and are an integral part of the world.
The session on expert systems included presentations describing a computer
simulation of perception of musical rhythm based on Lerdahl and Jackendorff's
generative theory of tonal music; a forward-chaining rule-based system that
performs harmonic chord function analysis for tonal music; a system for tonal
composition that learns by reordering its production rule priorities and by
generalizing new rules based on recurring patterns in its histories; and a
"Cybernetic Composer" that uses constraint satisfaction and backtracking to
compose realistic music in four different genre's ('standard' jazz, Latin jazz,
rock, and ragtime). During this last presentation, Charles Ames (Kurzweil
Foundation Automated Composition Project) played a wonderfully entertaining
audio tape of his system performing.
The session on tutoring systems and languages included presentations about
an architecture for a tutoring system to teach musical structure analysis and
melody interpretation; a tutoring system to teach elementary aspects of music
theory and ear training; a method of generating music using linguistic
(syntactic) techniques; a preliminary description of a programming language for
generating and analyzing musical compositions; and a knowledge representation
for tutoring systems using constraint satisfaction embedded in a logic
programming language (applied to teach harmony).
The session on cognitive models and knowledge representation had
presentations on modeling and generating music using multiple viewpoints based
on Markov models; a representation that provides sufficient expressiveness for
analysis of atonal music; and musical composition considered as problem
reduction.
The session on neural networks and parallelism included papers on storing
and processing time-varying musical information in PDP networks; learning of
musical concepts (keys, major/minor, scales, chord expectancies) and cognitive
properties (pitch invariance) using neural nets; and a method of applying the
massively parallel Connection Machine to separate audio sources in polyphonic
music. As part of this last presentation, Barry Vercoe (MIT Media Lab) showed
a videotape of an automatic accompanist that learns from rehearsal and adapts
to a performer's habitual idiosyncracies.
The final, rather eclectic session heard papers on the cross-fertilization
between the fields of music and AI; computer implementations of a cognitive
model of music perception; the use of spatial/visual representations and
improved user interfaces to aid sound analysis; and the need to realize the
limitations of computable functions as descriptions/simulations of music
cognition.
Len Moskowitz
Allied-Signal Aerospace
moskowitz@bendix.com (CSnet) Test Systems Division
moskowitz%bendix.com@relay.cs.net (ARPAnet) Mail code 4/8
arpa!relay.cs.net!bendix.com!moskowitz (uucp) Teterboro, NJ 07608