Panel discussion on neural stuff. Jan is speaking about self organizing maps, which is a talk he gave at brum last term.
He's making snapshots for presets. It can be used to find similar presets. That are like ones he likes.
It creates a meta controller, which is more high level.
He can use it to make sound objects.
And it's bewtween top down and bottom up appraoches.
He's got a graph on how he uses it. He plays with it to make snapshots. The snapshots are fed into the som which generates simlar material, which he can use for a meta controller. He can make a map of material amd then make a path to traverse the map andthen control where he is on the the path with a slider.
Soms can be used to control anything, including each other.
A snapshot can be an array or an event. His examples use ron's preset library.
it is an unsupervised neural network.
SynthDescLib lets you make a gui with preset. Or maybe this is jan's code lib. There is a button to generate a som from presets in the gui. And a matrix comes up. Some of them are green, which are the ones he picked. The others are related. As you click on them it saves your path. There is a slider at the top that moves through the path. You can save your state.
now dan stowell is recapping and he has made soms as a ugen. He is showing the thng he did at the london sc meetup. It runs on the server and gets trained in advance by ana;yzing samples.
it imposes the eq of one sample onto another sample. Which works and is impressive. The som has a visua;izer. Pretty. It is not for download. I find his gui is set up kind of in reverse of how i'd think about it.
too much coffee for me. Pee break now. Ok back.
now nick collins is showing his work on the topic. He's got an som implementation too. With a helpfile. He analyzes midi files. He breaks them up into little bits. He will release his files shortly.
now he's talking of reenforcement learning, which is a way of considering an agent acting in the world. (See david's photo of the slide) a state leads to an action, which in turn effects tje world which changes the state. Reenforcement learning looks at how effective actions are. So the program must have an idea of the world. This must also have a way of grading the reward of how good the world is. So you need to decide if something sounds good.
he has sc code to deal with this. LGDsarsa is on his website.
because machine learning is computationally expensive, it's often farmed out to an external batch process. Or you can run in non rt mode. Dan has a nice ugen for this called Logger. Thete's code examples on the mailing list. It creates data files which can then be used for machine learning.
he's got a self similarities table for a pixies track.
ok, on to the more panelly part
how do you get the reward state in sarsa? Physiological monitoring is one way. Or you can ask the audience which has a delay, but propgate backwards. Or you can do it in a model.
jan is doing a project with thom which is similar but will generate full pieces. Nick reccomends tom mitchells book on machine learning.
why is a reward better than a rule? Why is it more interesting to train a net vs creating rules? Answer is that they can be used for different applications. Ron notes that rules are implicitly present in selection of training material and assumptions. Nick is talking about flexibility and creative machines. Dan says that ron is correct, but the number of possibi;itites in even a small data set is huge. Ron says constraints are cool. Te panel says that supercollider is cool
james, or leader, is talking about intent. What if we inverted rewards to make the audience unhappy? Nick points out it's still hard to gauge cultural preferences.
there's a question about specificity vs building an overly large tool. Jan agrees this is a trap. Nick says that specificity is more musically effective. He talks about hard coding. There's too much variation sometimes.
performances with live evoluion. Using a human as a fitness fnction is slow. Nck talks abut a cmputer as an impovisor. His phd does this, whch you can download. He's switced to midi becaus featurewards extraction is hard. Jan s talking abut having few slders.