THE DISCIPLINE FORMERLY KNOWN AS MIR
Perfecto Herrera, Joan Serrà, Cyril Laurier, Enric Guaus, Emilia Gómez, Xavier Serra
Music Technology Group, Universitat Pompeu Fabra, Barcelona, Spain
{perfecto.herrera, joan.serraj, cyril.laurier, enric.guaus, emilia.gomez, xavier.serra}@upf.edu
1. INTRODUCTION
Music Information Retrieval is a young multidisciplinary
endeavour [4]. Even though its origins can be traced back
to the 1960’s [14], we all probably agree on the capital
influence that the International Conference on Music
Information Retrieval, started in 2000 as symposium, has
exerted on the sense of belongingness to a research
community.

Our exploration is not a science-fiction essay. We do not
try to imagine how music will be conceptualized,
experienced and mediated by our yet-to-come research,
technological achievements and music gizmos.
Alternatively, we reflect on how the discipline should
evolve to become consolidated as such, in order it may get
an effective future instead of becoming, after a promising
start, just a “would-be” discipline.

Our vision addresses different aspects: the discipline’s
object of study, the employed methodologies, social and
cultural impacts (which are out of this long abstract
because of space restrictions), and we finish with some
(maybe) disturbing issues that could be taken as partial and
biased guidelines for future research.

2. OBJECT OF STUDY
First of all, let’s face the misrepresentation that the name
of our discipline currently has. From the 3 words used,
only the word “music” is totally satisfactory.
“Information” is debatable in the sense discussed by
Wiering [14]: most of the interesting information that has
to be exploited is a-referential (i.e., not about the music
itself but about its context, functions, connections,
emotions, audiovisual links, etc.). Instead of “music
information”, we should use “information about music”. In
fact, what most of the algorithms and systems developed
in our community target is “music content”, something
that can be predicated, in a logical sense, about the music
(in other words: knowledge). To conclude, the term
“retrieval” just identifies a very narrow topic targeted by
some researchers attending to ISMIR and other related
conferences. But there is life beyond retrieval, and broad
and better suited alternatives for a permanent name would
be “processing” or “interaction” or even broader,
“research”. Therefore, the discipline formerly known as
MIR is probably going to be better characterized as “music
content processing”, “interaction with music information”
or (more generally, and if changing the acronym MIR is
seen as detrimental) “music information research”1.

Under our conception of MIR (or whatever other better
name we finally find for it) the discipline should be asking
profound questions and providing answers about music,
about the content treasured in short excerpts, tracks,
collections, or datasets, and about the context where it can
be experienced. If the developments, devices, systems,
experiments and evaluations do not increase our
knowledge about music then the field is going to get more
and more sterile and it will become just a technology
incubator, working just for the sake of technology. We
have mixed feelings on the level of the current
accomplishment of this quest for understanding music. In
the ideal future we foresee a wealth of new knowledge
about music and all kinds of interactions with music
information, not just a bunch of clever algorithms and nice
user interfaces. We will only develop music understanding
systems by means of understanding music understanding.

If we consider the Stokes’ quadrants [13], our discipline is
currently in the so-called Edison’s quadrant (pure applied
research) but its future lies in Pasteur’s quadrant, which
corresponds to use-inspired basic research.
Complementary pure basic research (Bohr’s quadrant)
should be generated by computer science, cognitive
science, musicology, or any other of the consolidated
disciplines that converge in our field (source disciplines)2.
This conceptualization is also paralleled by the
consideration of three types of physicists (or scientists we
would add) made long time ago by Victor Weiskopf:
machine builders, experimentalists and theoreticians. For
the sound and safe future of the discipline we need more
theoreticians and experimentalists to team up with the
excellent machine builders that have cropped up in the

1 This name is precisely the one adopted by the recently created
International Society for Music Information Research.
2 It could be discussed if basic research can be generated in the MIR
field.

early years of MIR. In the future, our discipline will be
more permeable to researchers grown in the source
disciplines, as they can bring theories, experimental
methods, and normative data (i.e., data to be used as
ground truth) to be exploited in the problems related to
music understanding by humans and machines. The
transition from Edison’s to Pasteur’s will be possible only
when the above mentioned shift of focus and percolation
of human resources is more common than now.

Research is inspired by… Considerations of use?
No Yes
Yes
Pure
basic
research
(Bohr)
Use-inspired
basic
research
(Pasteur) Quests for fundamental
understanding?

No
Pure applied
research
(Edison)

Figure 1. Research quadrants from Stokes [13].

Another feature of our discipline in the future is that work
will be predominantly done on the upper rows of Table 1.
Understanding music understanding has been mostly
pursued bottom up, and this has led to evidencing the
existence of a semantic gap [2]. Top-down approaches,
whereby a general model based on rational analysis of
behaviour or on task knowledge is developed, may
overcome this drawback. Examples of fruitful models that
are yet to be adapted to the specificities of interacting with
music content are, for example, the Information Foraging
Theory [10], or Rational Models that connect choice and
preference [8].

3. METHODOLOGY
Scientific rigour is not only to be demonstrated in the
proper use of mathematical models. It starts with a proper
use and acknowledgment of the past research legacy. In
our future papers, we will rarely see omissions of basic
works and no bias to prefer referencing authors of certain
countries or regions will be detected3. With the current
and future bibliographical research tools, omissions will
only be attributable to bad practice, either voluntary or
because of lack of care. Student supervisors will be very
concerned on playing the role of heritage convey-belts.


3 This observation arises from our own activity as journal and conference
reviewers. It should be substantiated with data that, unfortunately, cannot
be made public in most of the cases, or that would require extensive
bibliometric analyses.
Generalization poses an important overhead on our
research and it has not been seriously taken in many
papers, even those published in our most influential
journals. The claims written, often being very ambitious,
should be accurately substantiated using the appropriate
type and amount of instances, the correct type of tests, and
a baseline system or theory to compare and judge the
claimed improvements or advantages.

Level Question Analysis elements
Rational
What problem is
solved?
Resources, state and
phase dynamics,
affordances,
constraints
Knowledge
What does the
system know?
Goals, preferences,
semantic descriptors
Algorithm
How does the system
do it?
Front-end, low-level
descriptors,
classifiers, similarity
metrics
Implementation
How does the system
physically do it?
System architecture,
hardware devices,
software functions
Table 1. Levels of explanation adapted from [10]

The future MIR researcher will be aware of experimental
design and statistics, way better than the current one. It can
be scary to count how many t-tests can be found in the
available literature, and how many of these were not
justified or not properly done (the same has to be said on
the election of evaluation measures). It is also scary to
count how many published papers do not include any
formal hypothesis test at all (see, for example [3] for some
guidelines on good practices). Multivariate statistics,
bootstrapping, or Bayesian models will become off-the-
shelf tools for the future MIR researcher. In addition, the
Cranfield evaluation model (i.e., TREC-like) will be
enhanced with alternatives considering cognitive,
interactive and relevance issues that are not captured by
precision and recall measures [1].

One maturity feature of a discipline is the existence of
meta-analysis studies. Meta-analysis is aimed to the
accumulation of evidence across different studies targeted
to the same problem [11]. There are several musical
problems (e.g., genre classification) where the amount of
different studies, with different music collections,
algorithms, descriptors, classifiers and parameters qualifies
them for some meta-analysis integrative effort. In the
future, researchers will use meta-analysis as another tool
of the trade.
Subjective evaluations will be accomplished following the
formal requirements of subject sampling, briefing,
debriefing and ethical respect that are followed in other
experimental disciplines. A subjective evaluation that is

done as a “quick and dirty” way to demonstrate that “users
prefer the developed system” will be dismissed. As an
alternative, well-crafted experiments will be the norm,
were the effect of a variable (algorithm, system, interface,
etc.) is tested on controlled and well-defined indicators of
satisfaction, efficiency or task accomplishment.
4. THE FUTURE IS HERE; IT'S JUST NOT
WIDELY DISTRIBUTED YET
In the future, most of metadata will be generated in the
creation/production phase: This means that we mostly
need automatic content processing systems capable of
working with old content, and that content description at
the very moment of its generation should be more actively
facilitated by industrial-strength authoring tools.
We need less than 3000 semantic concepts to make
possible flexible description, visualization, retrieval,
navigation, etc: Is it possible to approximate, as video
retrieval experts did [5], this upper bound for music
content semantic description? Which could be the
consequences of that?
Everything will be done with social and human
computation: Even though automated systems are capable
of breaking yet-to-be-discovered glass-ceilings, the most
efficient and effective content descriptions are currently
(and will be in the future) created by humans. Instead of
putting all the effort on the automated analysis let’s try to
help humans to help other humans.
Music like water? Music as dog!!! Contrasting with the
fruitful and dominant metaphor of “Music like water” [6],
North and Heartgraves [9] conclude: ‘the [pop music]
audience places less emphasis on revering the music as
high art and more on it as a “friend” that supports them
throughout their everyday life, so that pop is less like a
god and more like a dog’ (p. 358). Do we want to develop
research and technology to support this metaphor or,
alternatively, to change it?
A digital economy has to be grounded on uncopiable
untangibles: Some of them are “personalization”,
“interpretation” (i.e., helping to enrich, understand or
complete the musical context of a work), “embodiment”
(e.g., how does it feel as attending a Pink Floyd concert by
1980?), “findability” or “community”4. The kernel of the
problems about interacting with music content is to be
found there, and our theoretical models are still far from
explaining and predicting music preference, navigation
patterns through collections of musical data or the
dynamics of music-related emotions.
4 http://www.kk.org/thetechnium/archives/2008/01/better_than_fre.php

Dynamics! Dynamics! Time information has been usually
given a secondary role in most of the developments of our
discipline (although, in other occasions, Hidden Markov
Models have been invoked for problems where more
parsimonious and simple solutions were available!). The
temporal dimension has to be moved to the foreground
also when looking at music-related communities: we do
not know anything about how they appear, grow and
influence the existing ones. Our listeners and users have
also been modelled as finite-state machines and
approaches based on dynamical systems are still rare [12].

5. REFERENCES
[1] Borlund, P. The IIR evaluation model: a framework for evaluation
of interactive information retrieval systems. Information Research,
8 (3), paper no. 152, 2003.
[2] Celma, O., Herrera, P., Serra, X. Bridging the music semantic gap,
1st International conference on Semantics And digital Media
Technology (SAMT), 2006.
[3] Dietterich, T. G. Approximate statistical tests for comparing
supervised classification learning algorithms. Neural Computation,
10 (7) 1895-1924, 1998.
[4] Futrelle, J., Downie, J.S. Interdisciplinary communities and
research issues in Music Information Retrieval. Proceedings
ISMIR 2002, 215–221, 2002,
[5] Hauptmann, A., Yan, R., Lin, W. How many high-level concepts
will fill the semantic gap in news video retrieval? CIVR 2007
ACM International Conference on Image and Video Retrieval,
Amsterdam, The Netherlands, 2007.
[6] Kusek, D. and Leonhard, G. The Future of Music: Manifesto for
the Digital Music Revolution, Berklee Press, 2005.
[7] Law, E. The problem of accuracy as an evaluation criterion. ICML
Workshop. on Evaluation Methods in Machine Learning, 2008.
[8] Lucas, C., Griffiths, T.L., Xu, F., & Fawcett, C. A rational model
of preference learning and choice prediction by children. Advances
in Neural Information Processing Systems 21, 2009.
[9] North, A., and Hargreaves, D.J. The Social and Applied
Psychology of Music. Oxford, Oxford University Press, 2008.
[10] Pirolli, P. Information Foraging Theory: Adaptive Interaction with
Information. Oxford, Oxford University Press, 2007.
[11] Rosenthal, R. and Dimatteo, M.R. Meta-analysis. In Pashler, H.
(Ed.) Stevens’ Handbook of Experimental Psychology, Third
Edition. Volume 4: Methodology in Experimental Psychology.
New York: Wiley, 2002.
[12] Spivey, M. The Continuity of Mind. Oxford, Oxford University
Press, 2008.
[13] Stokes, D.E. Pasteur’s quadrant: basic science and technological
innovation. Brookings Institution Press, Washington, D.C, 1997.
[14] Wiering, F. Can Humans Benefit from Music Information
Retrieval? In Marchand-Maillet, S., Bruno, E., Nürnberger, A.,
Detyniecky, M. (Eds.) Adaptive Multimedia Retrieval: User,
Context, and Feedback, Berlin, Springer, pp. 82-94, 2007.