On the spectral bias of deep neural networks