I understand what ICA does at a high level but in the cocktail party problem context. All the examples, articles I have read take a similar problem to explain ICA where the aim is to derive the independent sources from the mixed readings recorded.
But I was introduced to ICA in the dimensionality reduction context where I studied PCA first and then moved to ICA.
I am confused as to what ICA is really giving out when comparing its output to PCA. I was reading an article here which quotes the following:
ICA is an algorithm that finds directions in the feature space corresponding to projections with high non-Gaussianity. These directions need not be orthogonal in the original feature space, but they are orthogonal in the whitened feature space, in which all directions correspond to the same variance.
Could anyone please explain :
- The "high non-Gaussianity" thing in ICA
- What is whitening in ICA?
- Summary of what the above quote is really conveying
Using python's scikit, I plotted the graph for PCA and ICA on the dataset I am working with. Here is what they look like after reducing to 2 components:
Could we conclude/comment something on the data just by seeing the PCA / ICA plots?