Learning with Transformation Invariant Kernels