I’ve had this idea for a while, but was constantly debating in my mind whether it was the same thing as an SVM with a kernel. I finally came to the conclusion that it wasn’t:

Start with a hyperplane boundary and then split it into warp/deformation points. Optimize the warping by minimizing MSE using gradient descent (or something) and it should begin to take on the form of the points. Impose some sort of regularization to prevent overfitting. Since warping can modify the plane in three dimensions, it’s no longer a hyperplane, but neither is it equivalent to passing a hyperplane through a kernel.

It should be a very powerful means of performing regression (and thus classification). I might research it later; I have too much on my plate now.