While trying to learn more about these concepts which are entirely new to me, I came across this website. This is not for our ‘Masters’, rather might be a helpful resource for the laymen.
This is indeed very interesting, but I would not call that a laymen text! It takes someone who wants to understand under the hood indeed, but with the graphics it is true that one could read it and ignore the Python part and get the take-home message. I’m sure @weefuzzy will be happy to know about this resource. It is indeed very much in line with @groma 's paper/experiment/demo, and I’m happy it helped you understand the variations between data-reduction algorithms, and gains to be had in comparing them in a single process.
Soon our code will be out on github with the paper, but it is more exploratory than gig-ready. Research code, as I’m told. But that is good for exploring these differences indeed with your own corpora.
Jacob told me that this kind of stuff will be part of the tool set, which is great. And I regret missing out on the sessions two weeks ago, as I was abroad. In any case, I’m working on some stuff for which I’m interested in this kind of thing. So thanks @tutschku for sharing this! I especially like the plot of the T-SNE algorithm. But maybe that’s just a dumb aesthetics informed response. Which will be implemented? And when will it be available?
(Apologies if these are obvious questions.)
Next week, @groma will present the SuperCollider environment that allows to do some comparative tests between various data reduction algorithms - the code should be public then, so we’ll post it here.
This is a great question, which was also asked at the plenary, and funnily was sort of the motivating of our research, but from the point of view of music practice we concluded there is not a single best algorithm, and it all depends on what you want. The paper has some better explanation and will be available in our website soon.
Another motivation was that, as the tutorial @tutschku links shows, scikits-learn has a quite nice collection of implemented dimensionality reduction algorithms, so we used that in our research, by running python from SuperCollider. Our aim is to make the project more useable by porting some algorithm to SuperCollider. This should happen soon, in the context of our NIME presentation. In toolbox two, some of these will be implemented as C++ objects for SC, Max, etc. This will take much ore time though, we will privately share an alpha version in November.