Machine learning: categorization of patches

[christofmuc]

I made an experiment, and it seems there is easily enough information in the sysex that when you have a learning data set (have manually or via name assigned categories to a couple of patches), you can train a classifier and then have the rest of the patches automatically assigned into these clusters.

First experiements on the Matrix 1000 showed this works. Kind of.

Not sure about multi-assignments - a cluster algorithm always will choose just one class, but we can actually deal with uncertainty.

Main challenge to productization will be that your python environment will need to contain scikit-learn and other huge python libraries, which is not a friendly install for everybody.

[markusschloesser]

what did you have in mind w re to which clusters? Also wouldn't the ML need to be aware of the structure of the sysex files?
Sounds really interesting!
One thing which comes to mind, especially when taking your "search for specific parameters" issue into consideration, is searching for sounds with a short decay and short release and putting them under "plucky". Or long attack, long release = pad

[christofmuc]

Ah! Good one - I could extend the auto-classification from currently using only regex matching on the name of the patch to full calculated predicate expressions on the parameters. So this would be an extension on #51.

Regarding the sysex structure - maybe. My tests ran on extracted parameter values for the Matrix 1000, so yes, it could help. But on the other hands, the learning/clustering algorithm are so generic, you could also just pipe in the sysex vector and see what happens. When you do princpal discriminant analysis, many values will be disregarded anyway to distinguish the clusters. I think its worth a try.