CODEWITHSUNDEEP
pcadimensionality-reduction
starrr
starrr

Reputation: 1023

When using ICA rather than PCA?

I know that PCA and ICA both are used for dimensionality reduction and in PCA principal components are orthogonal (not necessarily independent) but in ICA they are independent. Can anybody please clarify when it is better to use ICA rather than PCA?

Upvotes: 4

Views: 1854

Answers (1)

lejlot
lejlot

Reputation: 66795

ICA is not a dimensionality reduction technique. ICA is used for separation of convolved signals, which might have smaller dimension than the input space, but this is rather a side product, not aim as such. Thus ICA and PCA have different fields of applications. PCA is usually used to visualize high-dimensional data (through selection of 2 principal components) or simply to reduce the dimension to the one one can handle with given method. ICA on the other hand is used when you have signals, which are heavily convolved and you want to separate them, thing for example about two people speaking in the same room, recorded with 2 microphones. ICA will be able to separate each speaker, while PCA would fail. Similarly, ICA will look for non-gaussian, convolved signals, thus if your data is at least to some reasonable extent - gaussian in nature, it will destroy this structure (as the underlying assumption is that this is not true).

Upvotes: 10

Related Questions