Independent Component Analysis (ICA)

[pfeatherstone]

Is there a way to do ICA in dlib using some of the traditional machine learning tools? Something that mirrors https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.FastICA.html

[pfeatherstone]

Specifically, I'm trying to write an algorithm with dlib which extract pulse rate from a video. The algorithms typically do:

• detect face
• average RGB pixels from face excluding eyes, mouth, hair
• track (which gives you R(t), G(t), B(t)
• Perform ICA on (R(t), G(t), B(t)) which gives you (X1(t), X2(t), X3(t)) which are 3 linearly independent channels
• Perform FFT on X(t) and take max peak
• Peak gives you pulse rate

Something like that. The bit I'm missing is ICA.

[arrufat]

It only compresses if you don't use all dimensions. If you use all dimensions, there's no compression.

[arrufat]

Basically, it will put an axis on the dimension that has the largest variance (or energy) and start appending dimensions orthogonally to the previous dimensions from highest to lowest variance.

[davisking]

Na you need to code up something for ICA. The linear algebra tooling should be useful. I don't think the solvers or kernel stuff would help you though.

[pfeatherstone]

I don't know enough about traditional machine learning stuff. This gives a good comparison of PCA vs ICA.

[arrufat]

Yes, it seems like PCA is not doing what you need. It's not suitable for this problem, since the sources are not orthogonal, it can't separate them.

[pfeatherstone]

It feels like dlib already has all the necessary tools to do this. E.g. using the kernel solvers

[pfeatherstone]

Maybe linearly_independent_subset_finder is relevant here. Not sure. I'll read the examples but this isn't my area of expertise.

[pfeatherstone]

https://en.wikipedia.org/wiki/FastICA spells the algorithm out. Could use this over trying to copy-paste sklearn or some other implementation. Hopefully, the unit tests would show near identical results when using dlib or sklearn on same input data.

[pfeatherstone]

We could also eventually add ILRMA, FastMNMF and other blind source separation algorithms which are related to ICA. I've used FastMNMF in the past for solving cocktail party problems but never thought more generally of other use cases.