5 ms·

I don’t think a computer could know what the artist intended to create. I wonder, how accurate can the computer be at predicting the next note (the first note

by kovek 5y ago

I don’t think a computer could know what the artist intended to create. I wonder, how accurate can the computer be at predicting the next note (the first note from the uncompleted section)? Is that an interesting question?

kace91 5y ago
That can be proven though. Keep a work by the author out of the training phase, then see if the model can predict it from an incomplete version. Otherwise...
6gvONxR4sf7o 5y ago
Yes, it's a very interesting question! This how many language models work. You might be interested in reading about a metric called perplexity. [0] The major caveat is that it's invalid to evaluate on completed, polished work, and then extrapolate to what's practically a napkin note. Another important note about it is that some next token predictions are harder than others. For example, try predicting the next word (the _____) in each of these: > I really love this Dr Seuss Book called "The Cat in the _____ versus > I really love this Dr Seuss book called "The Cat in the Hat." _______ [0] https://en.wikipedia.org/wiki/Perplexity https://en.wikipedia.org/wiki/Perplexity
- noduerme 5y ago
  You could also make a model which always responds with the word "hat". If the second sentence is the "creative" portion, it would then prove 100% accurate on your in-sample data, and the next sentence could be "Hat, hat hat." And then people could say, see, this brilliant AI speaks like Dr Seuss! You could also just string a bunch of random words together and build a model that learned to predict them with total accuracy. Of course if the words are truly random, it would have basically no chance of predicting the next random word from outside the learning set. The conceit really is that some model got enough important, critical contextual information from the partial symphony to be able to predict with greater than random probability the next portion. On a micro level, one note at a time, maybe. But every note further into the unknown is predicated on the previous ones, so on a macro level it's absurd to think it's "predicting" anything.
- diskzero 5y ago
  Thanks for sharing that term. A score is so much more complex. I need to dive into the approach Playform took. I notice that team is comprised of experts in a variety of fields, who make a lot of editorial decisions on the output of the AI and then decided the style of the output (scherzo, etc.) and where that particular element might be placed in the score. I wonder what would have happened if they just ignored the AI and completed it using their own compositional knowledge?
diskzero 5y ago
Calling the prediction accuracy doesn't seem quite right, but I don't know what term I would use. I created a mobile music notation application that has pretty good accuracy when writing on the score. Some users asked for the ability to predict what might be written next. The first challenge is creating a UI to present the choices, but lets assume that is a solved problem. I trained up models and you can probably guess what happens. It is similar to the demos you have listened to where thirty seconds of on original song are then taken over by AI. Things sound similar and then descend into madness. This is sort of fun but not super useful to a composer. Maybe if you, the composer, had a large corpus used to trained with, the AI could riff on your own ideas. My experiments were with a single instrument and then I tried using the process on multiple instruments. I would consider the end product pretty much useless. Users of music notation software, especially those who are able to write out their compositions using notation, would find this feature pretty aggravating. Cool for bizarre demos, pretty useless for making a living. I do expect to see a lot more activity in this space as we learn more techniques, train bigger models and learn how to keep the AI from going off the rails.

by kovek 5y ago

kace91 5y ago

That can be proven though. Keep a work by the author out of the training phase, then see if the model can predict it from an incomplete version. Otherwise...

6gvONxR4sf7o 5y ago

Yes, it's a very interesting question! This how many language models work. You might be interested in reading about a metric called perplexity. [0] The major caveat is that it's invalid to evaluate on completed, polished work, and then extrapolate to what's practically a napkin note. Another important note about it is that some next token predictions are harder than others. For example, try predicting the next word (the _____) in each of these: > I really love this Dr Seuss Book called "The Cat in the _____ versus > I really love this Dr Seuss book called "The Cat in the Hat." _______ [0] https://en.wikipedia.org/wiki/Perplexity https://en.wikipedia.org/wiki/Perplexity

noduerme 5y ago
You could also make a model which always responds with the word "hat". If the second sentence is the "creative" portion, it would then prove 100% accurate on your in-sample data, and the next sentence could be "Hat, hat hat." And then people could say, see, this brilliant AI speaks like Dr Seuss! You could also just string a bunch of random words together and build a model that learned to predict them with total accuracy. Of course if the words are truly random, it would have basically no chance of predicting the next random word from outside the learning set. The conceit really is that some model got enough important, critical contextual information from the partial symphony to be able to predict with greater than random probability the next portion. On a micro level, one note at a time, maybe. But every note further into the unknown is predicated on the previous ones, so on a macro level it's absurd to think it's "predicting" anything.
diskzero 5y ago
Thanks for sharing that term. A score is so much more complex. I need to dive into the approach Playform took. I notice that team is comprised of experts in a variety of fields, who make a lot of editorial decisions on the output of the AI and then decided the style of the output (scherzo, etc.) and where that particular element might be placed in the score. I wonder what would have happened if they just ignored the AI and completed it using their own compositional knowledge?

diskzero 5y ago

Calling the prediction accuracy doesn't seem quite right, but I don't know what term I would use. I created a mobile music notation application that has pretty good accuracy when writing on the score. Some users asked for the ability to predict what might be written next. The first challenge is creating a UI to present the choices, but lets assume that is a solved problem. I trained up models and you can probably guess what happens. It is similar to the demos you have listened to where thirty seconds of on original song are then taken over by AI. Things sound similar and then descend into madness. This is sort of fun but not super useful to a composer. Maybe if you, the composer, had a large corpus used to trained with, the AI could riff on your own ideas. My experiments were with a single instrument and then I tried using the process on multiple instruments. I would consider the end product pretty much useless. Users of music notation software, especially those who are able to write out their compositions using notation, would find this feature pretty aggravating. Cool for bizarre demos, pretty useless for making a living. I do expect to see a lot more activity in this space as we learn more techniques, train bigger models and learn how to keep the AI from going off the rails.