1

Currently I'm working on facial recognition. If I use encoding/feature vectors of 2 images which method will prove more accuracy, L2 norm or cosine similarity and why?

I read "ICA performs significantly better using cosines rather than Euclidean distance as the similarity measure, whereas PCA performs the same for both."

Why is this true? When those methods will fail? And if other better approach is possible.

offset-null1
  • 55
  • 4
  • Cosine similarity converted by the [cosine rule](https://stats.stackexchange.com/a/36158/3277) into a distance is called chord distance which is a case of euclidean distance. Therefore, analysis based on cosine is most of the time equivalent to the analysis based on squared chord distance. – ttnphns Sep 22 '20 at 12:03
  • Actually I have heard that cosine similarity gives better results but I'm not clear with the reasons. – offset-null1 Sep 22 '20 at 12:07
  • You should clearly express what is PCA or ICA performing on _a distance_ (rather than a dataset or a covariance matrix). – ttnphns Sep 22 '20 at 12:48
  • Do you mean what's their functionality in this case?If so ICA is for extracting independent basis and pca compresses the matix – offset-null1 Sep 22 '20 at 13:18

0 Answers0