NORMAS DE OBLIGADO CUMPLIMIENTO: EDIFICACION
6. Fases de Ejecución 1. Demoliciones
6.3. Trabajos Previos
What, then, is the “theory of musical meaning” employed by Spotify? Above I have sketched part of an answer: that music’s meaning is functional rather than intrinsic, and that the mysterious ways in which music causes people to feel things – whatever they are (and Whitman definitely doesn’t try to answer that) – will necessarily appear in a meaningful way somewhere, provided we gather enough data. In short, the “theory” of musical meaning is nothing more than an expression of the assumptions grounding the fields of machine learning and pattern recognition.
The fact that Whitman’s system is able use words like “learn” and “meaning” while essentially skirting the kinds of thinking that traditionally attach to them is part of a broader trend. In the rhetoric of machine learning, it is almost as if with enough data, theory itself becomes obsolete.61 Data, if the quantity and quality are high enough, can seem to supplant
61. Chris Anderson, “The End of Theory: the Data Deluge Makes the Scientific Method Obsolete,” 2008, accessed July 25, 2019, https://www.wired.com/2008/06/pb-theory/.
theory. Predicting supplants understanding and explaining. That, at least, is what is happening here with respect to music. As Zarsky puts it, quoted above, this is the assumption:
That human conduct is consistent and that with sufficient data human behavior becomes predictable.62
Turning back to the musical domain, though, is that a “theory” at all? You might answer “no,”
and you might be right. But what, then, do we make of Whitman’s claim to have “learned the meaning of music?” If we do not have a notion in place of what musical meaning is, what is it exactly that we have learned? And what, then, do we make of Spotify’s claim to be worth $10 a month? Are not Whitman’s dissertation and Spotify’s pitch equally grounded, on some level, on the idea that they’re at some level right about what music means? And is being basically right about musical meaning not ipso facto a kind of theorizing?
Whether we approach recommendation through Spotify’s GUI or Whitman’s dissertation, it seems clear that some kind of theory is at work; there’s no other way to “learn the meaning of music.” And yet the theories there, such as they are, have no real explanatory power; they’re theories of data aggregation and prediction. The recommendation problem, then, resumes the familiar problem from the critique of AI in culture: whether machine intelligence can be evaluated behavioralistically or not. It resumes it, but it doesn’t do anything to resolve it. This can be seen as a flaw in its design, or it can be a simple expression of the fact that it’s a product, not a piece of knowledge. In other words, there’s a parry available to Spotify’s apologists here: very well, you might say, Spotify is wrong about meaning. So what? It’s not a form of scientific knowledge, but just a collection of engineers trying to solve a problem and earn some money. But, as Pelillo et al (2015) argue, the era of machine learning has changed the way we should think about this traditional distinction:
The scientist’s occupation is seen today more modestly as a kind of problem-solving activity not dissimilar conceptually to that of the engineer, whereas on the other hand
62. Zarsky, “The Trouble with Algorithmic Decisions: An Analytic Road Map to Examine Efficiency and Fairness in Automated and Opaque Decision Making.”
the work of the engineer is thought to produce a form of knowledge which is on a par with that produced by the scientist.63
Whitman himself echoes this idea; without giving a full throated endorsement of “strong AI,” he does seem to suggest that insights about man and insights about machine don’t necessary stand at odds with each other:
You call it algorithms but it’s a lot more than that. We are obviously doing a ton of computer stuff but it’s all based on what people are saying and choosing and that stuff. We hate this stupid man versus machine dichotomy.64
If the man-vs-machine dichotomy is stupid, it should follow that the programmatically derived “meaning” is not just an engineering expedient, but some kind of true statement about how music works. If Spotify is worth paying for, it should also follow that Whitman’s technology – culling meaning from aggregated human activity and linking that to the audio signal – should have really captured musical meaning. It’s implicit in saying that you’ve learned the meaning of something that you have some explanation of how it works, that you’ve cracked some kind of code. But when you actually look at Spotify’s recommendation engine, what you have is, on one hand, an expression of the reducibility of human conduct to statistical predicition (nothing more than an axiom underlying all of machine learning) and, on the other, a theoretically feeble
“relationality,” which isn’t even all that different from what we find in Leonard Meyer (1956).
Whitman sidesteps, in other words, the question that should matter to him most (what does music mean?), even as he postulates a cryptic kind of answer. Compounding matters is the fact that, in the case of Spotify, that cryptic answer is itself never divulged to the consumer in any way. Still, we can say one thing for certain about what that theory is: that if we collect enough data, musical meaning, whatever that may be, will inevitably, somehow, be captured by
63. Marcello Pelillo, Teresa Scantamburlo, and Viola Schiaffonati, “Pattern recognition between science and engineering: A red herring?,” Philosophical Aspects of Pattern Recognition, Pattern Recognition Letters 64 (2015):
3–10.
64. Emily White, “The Echo Nest CTO Brian Whitman on Spotify Deal, Man Vs. Machine, Why Pandora ’Freaks’
Him Out (Q&A)”, Billboard interview in 2014, https://www.billboard.com/biz/articles/news/digital-and-mobile/
5944950/the-echo-nest-cto-brian-whitman-on-spotify-deal-man-vs, accessed May 20 2019
the system. As for the nature of that meaning, Whitman takes a peculiar approach: he cites a single source (Meyer) as representative of countless humanistic authors who have thought about the issue, gives that source a cursory reading (it’s “absolutist”), and declares his own system to be an improvement upon it. To the criticism that the system tramples over the nuance of true musical experience, or to the tradition of aesthetic philosophy, Whitman will always have the ready defense that this doesn’t matter since that the real task is software design rather than philosophy.
Ignoring basic questions like this is a kind of privilege, the same one enjoyed by Roger Schank when he disavowed philosophizing about technology decades earlier (see Chapter Two).
The prestige and monetary value of the tools under development by Schank and Whitman drown out the simple questions we might want to ask about them: does this really stuff work? And, if so, how? That these questions are hard to hear over the din of excitement surrounding machine learning today does not mean that they shouldn’t matter to the software designer; what this chapter argues, more than anything else, is that they do matter.
A close look at Spotify’s treatment of the problem of musical meaning shows that the advent of big data approaches does not magically make that problem disappear; that even the one philosophical source Whitman cites is actually aware of it in very much the same terms as Whitman, and that, above all, it remains as obstinate a problem as it has been throughout its long history in aesthetic philosophy. It is as thorny a problem as the financial crisis confronting the music industry in the 21stcentury, to name another problem Spotify hasn’t solved.
. . .
Chapter Three, in part, is published in INSAM: The Journal of Contemporary Music, Art and Technology, No. 2, Vol. I, July 2019, pp. 36 - 64, as “What Does Music Mean to Spotify?
An Essay on Musical Meaning in the Era of Digital Curation.” The dissertation author was the sole investigator and author of this paper.