Dimensionality reduction for temporal organization

[tremblap]

This is epic, thanks! I’m on a train so I cannot watch the videos yet but I might tease you in the other thread r.e. your last paragraph here What I like very much is the transparency of the process, where data mining is investigative and put against itself.

A few thoughts right away, on the more ‘objective’ matters:

@tedmoore

3. I analyzed those slices and extracted audio descriptors (714 columns by the time the 1st derivative and all statistics were done!).
4. I reduced that to 11 principle components (and was able to retain 99% of the variance!).

#4 means that you had a lot of redundancy in #3. But @groma will confirm that what I say is true.

I’m impresse by how well UMAP has fared - I presume you tweaked the local-vs-global topology priorities a bit? Could you share how you manage to tweak it?

@tedmoore

2. I sliced all that audio into 100 ms slices.

were you not tempted to do a more perceptual segmentation? did you try and abandoned? I’m very curious about the investigative process, where you stalled and how you curated that clean list. Maybe it was just as is from the outset, but maybe not?