The network is the music

August 9, 2015 1:06 AM

Here's how to create and train your very own neural networks to compose classical music! Technical background, cool illustrations and music samples included.
posted by Foci for Analysis (14 comments total)

It's always so funny to me that these are called "neural" networks, since there aren't actual neurons involved, only their electronic analogue. In a future world of artificial persons which I know will never exist, I imagine them getting annoyed with us whenever we call transmitting nodes "neurons," and then lecturing us on the biases we're bringing to thinking, not unlike a Canadian might chafe mildly at the fact that their nation is in North America and yet their southerly neighbors keep insisting on calling themselves uniquely "American."
posted by koeselitz at 1:29 AM on August 9, 2015

That's because Canada is North of America.
posted by I-Write-Essays at 2:56 AM on August 9, 2015

I generated some classical music but it's full of dogs and eyeballs. Did I do something wrong?
posted by Mr.Encyclopedia at 5:47 AM on August 9, 2015 [10 favorites]

Mine turned off the fridge and started to sing Daisy, Daisy...
posted by Devonian at 5:50 AM on August 9, 2015 [2 favorites]

If you dig this, check out this piece that mxcollins recently posted to MetaFilter Music: Harmonia
posted by usonian at 8:39 AM on August 9, 2015

The assertion that the multi-layer perceptron is not really a network of neurons is a very old one and has been repeated many times by many distinguished folks and it doesn't stop the functionality from being damn good. When Rumelhart, Hinton and the PDP folks generally promulgated the backpropagation (or the other guys who claim the invention - whatever, it's just a fancy-ass way to use the chain rule, anyways) they sat down and thought quite seriously in dynamical systems terms to say that the phase space of the MLP was the thing that was simulating neural dynamics, and that was to be the minimal simulation of things. Do you really want to be simulating individual neurons' biochemistry? (... yes, say some of the biologists, but they have had less success at actually doing computations) And they thought hard to poke at universality, and things in that dynamical vein. So the claim was only ever about phase spaces and functionality.

I tend to think that recurrence is not sufficient for poking really at the kinds of data that humans are truly good at: necessary, maybe, but not sufficient. I think something that can sit down and do the inverse problem for a signal at criticality would be sufficient.
posted by curuinor at 10:11 AM on August 9, 2015 [1 favorite]

On doing the inverse problem for a signal at criticality: compare to Voss's and Clarke's claims about music. Note that the PDP-11 in the paper is definitely different from the PDP group I was talking about
posted by curuinor at 10:14 AM on August 9, 2015 [2 favorites]

One of the things you'll hear in the machine learning community is that we can expect neural networks to catch up to humans on any activity that we can perform in a fraction of a second. This includes things like facial recognition, and, when you think about it, musical improvisation, which we're seeing here. What the net lacks is large-scale structure, which comes in part from longer-term memory. The way memory is implemented in these RNN's makes it quite difficult to have memories which last more than a few seconds; finding ways to build longer-term memory and use it interesting ways would be fantastic. Thinking of musical structure, we hear in these pieces lots of interesting snippets, but without any overarching structure or themes; one wants an algorithm that generates interesting themes, mutates the themes, and transitions gracefully between them, while still displaying the short-scale virtuosity if this algorithm.
posted by kaibutsu at 10:33 AM on August 9, 2015 [3 favorites]

Yet another impressive accomplishment in computer programming for a task that isn't really needed, but is god-like in its purpose.
This machine written music sounds competent but without soul or any "hook" as modern songwriters call it.
It would fit in well on a sirius classical channel--you know, where an algorithm builds the playlist.
posted by Fupped Duck at 10:45 AM on August 9, 2015

Because there aren't enough talented people out there who'd love to have an audience for their music.

Fascinating question is why -- in a world where life isn't as cheap as it once was ... but there are still plenty of people who can do stuff that no machine will ever be able to do in our lifetimes ... that are almost always barely tolerated -- we have been fascinated with automatons for a thousand years. Clearly the fascination is not about the value of the performance.
posted by Twang at 11:21 AM on August 9, 2015

kaibutsu: That's what criticality is for! That's the real principled path to true long-term memory, at least in my opinion, not the neural Turing machine stuff the deepmind people have done, cool as it is.

If I was a crank, I would be among the million cranks this sort of field turns out pretty regularly, as a matter of fact.
posted by curuinor at 12:16 PM on August 9, 2015

This machine written music sounds competent but without soul or any "hook" as modern songwriters call it. It would fit in well on a sirius classical channel--you know, where an algorithm builds the playlist.

Well, I just skimmed the implementation notes, so I might be missing something—but it seems that the compositions are being played to a strict metronome, with no variations in dynamics (i.e., every note is equally loud), and using a pretty crappy sampled piano patch. (High-quality virtual instruments for piano are expensive—creating them requires a lot of painstaking sampling.)

So, you can't really compare these pieces with a conventional classical performance. It would be fairer to compare them to a (human-written) classical composition played back with the same MIDI software—so that both are subject to the same limitations. (Or, conversely, you could get a human pianist to learn the neural-network-generated scores, and give a more humanized performance of them.)

There's probably no reason that this technique couldn't be extended to account for dynamic changes, variations in tempo, etc. I have no doubt that we'll eventually see software that can generate passable music in a variety of styles. It probably won't ever replace human music—but it might find a role as Muzak, soundtracks for video games (a few games already feature algorithmic music), a ready source of original-ish music for more functional (as opposed to artistic) purposes such as commercials and video production, etc.
posted by escape from the potato planet at 7:30 PM on August 9, 2015

but it might find a role as Muzak, soundtracks for video games (a few games already feature algorithmic music), a ready source of original-ish music for more functional (as opposed to artistic) purposes such as commercials and video production, etc.

You may be right on the technology, but I don't know why it would ever be used, from an economic perspective. Music, in all forms, is subject to massive oversupply. There are many more competent composers and performers who wish they could make a living from their craft than we will ever have jobs for.
posted by howfar at 1:14 PM on August 10, 2015

howfar:
Imagine a game which created music procedurally in reaction to your recent gameplay. Like your own personal soundtrack, responding to what you've been up to. There's a huge range of possible action, especially in an open-world game like Minecraft, and having an engine that creates music to order would be pretty cool.

The crazy part of automation is that it opens up possibilities we wouldn't have even considered before, because they were too far outside the realm of the possible to even register.
posted by kaibutsu at 12:13 AM on August 11, 2015