Variable selection with false discovery rate control in deep neural networks