Estimating Equivalent Kernels for Neural Networks: A Data Perturbation Approach