Class "interpolation" (distance to classes)

[weefuzzy]

@rodrigo.constanzo

So as a reminder, the general idea is that (based on a great suggestion/chat with @balintlaczko) each class would be one-hot encoded and then fed into fluid.knnregressor~ where the amount of nearest neighbours would be be used to determine how much of that class it was. If all the neighbors were the first class, that meant it was 100% that class, if it was 50/50, that would be 50% etc…

My thinking/hope was that using outlier rejection to create a more condensed and representative “class” would therefore make the one-hot encoding more accurate, therefore when moving from one class to the other, that that would create a smoother overall change.

As a general note, I think this is an optimistic approach given the challenging material (very short analyses of onsets) and one-hot encoded classifier output. Me and Rod have been talking about his data offline for a couple of days whilst I prod it in python as a case study for what sorts of feature and model selection / evaluation stuff it might be useful to add to flucoma in the future. One possible takeaway is that having the option of softmax rather than one-hot encoding might be more useful for this (but also, I think trying to get a regression to give a reliable linear-ish feeling of between class-ness for this stuff still might be ambitious, not least because it’s asking for more sensitivity to difficult data than the original classification).

@rodrigo.constanzo

It appears that there is a limit to how smooth this approach can get with one-hot encoded nearest neighbours.

Indeed

@rodrigo.constanzo

In terms of processing datasets. I do feel like some way to split/iterate/pack datasets (and labelsets) would be very helpful.

Yes. I have some Max abstractions brewing, which would at least allow us to try out some interface ideas in Max-ish environments. Filtering a dataset by associated labels isn’t too horrible, but also not the sort of code you want to have to write repeatedly. Maybe we can start a new thread and work out what a usefully generic collection of utilities for this would look like.

@rodrigo.constanzo

Is there a way to implement the outlier rejection such that once computed/fitted, a new incoming data point can be checked against both classes in a way that would offer a smooth(er than I get with one-hot encoding) “interpolation” of where you are relative to both classes?

Seems like a stretch. It can be augmented to score incoming new points (against the fitted data) but the algo isn’t holding enough information about how neighbourhoods of points relate to each other to be able to help much with determining whether a point lies near A, not to far from B but miles from C. That still feels like a straightforwardly supervised regression problem.