Edited by author 09-27-2001 02:51 AM
1. I think the most practical function of PCA is to do feature extract and dimention reduction, not noise reduction, while KPCA has a main advantage that it can capture the data structure easily, so that the noise that deviate the main structure can be percepted. So I don't think it is proper to compare the denoising result between linear PCA and KPCA. I am wondering how about the result if comparing the denoising experiment of kernel PCA and other classical denoising methods.

2. I think there is still some assumption about the data model: PCA can't deal with the uniform data. But practically, in real world, the data wouldn't be uniform. So that's why always we can adopt PCA without any concern.

3. If we only use the kernels that always have pre-image, then Gaussian kernel would be banned. A great loss!