On the Spectral Bias of Deep Neural Networks