On the rates of convergence for learning with convolutional neural networks