How to know which Kernel is better?

Question

I am working on an Image recognition software -

My first question is since I already explicitly turm my training images to features vector (and also my test images) what is the point of using Kernels to begin with? since, from what i understand, they are supposed to "save me the trouble" of explicitly computing the feature map for every point in the trainint set.

Second question: Given a specific feature map (for instance in image recognition: size, color, texture, brightness etc) How can I know if I should use a Linear Kernel, Polynomal Kernel, RBF or maybe some custome kernel? and if custome is the answer how to design it?

Answer 1

Kernels don't save you the trouble of vectorizing images. In fact, no learning method does. Of course, you can write a wrapper to do this once and be done with it. Kernels do often save you the trouble of identifying a suitable feature space, through the use of the kernel trick.

You can find out which kernel function is best in two ways: (i) try different kernels and see what works and (ii) read about what comparable applications have used. The former is most reliable by far, and can be automated using hyperparameter optimization libraries like Optunity.

If you want to obtain good representations and learn distinctive features automatically you should have a look at deep learning.