How does high dimensionality impact deep learning generalization?

It increases risk of overfitting but deep models exploit structure to generalize.

Despite over-parameterization, deep networks generalize due to implicit regularization, data structure, and optimization biases like flat minima preference.