Laplacian Regularized Few-Shot Learning