Interesting work. A DNN is after all something that compresses the input data and hopefully generalizes well over the domain. This 'machine' itself will adhere

Interesting work. A DNN is after all something that compresses the input data and hopefully generalizes well over the domain. This 'machine' itself will adhere to principles of Kolmogorov complexity etc.. Traditionally, empirical solutions in the NN literature to address instability are regularization and dropout. Also, adding layers seems to improve things. The famous example is XOR which cannot be trained by a single layer NN. How do the theoretical limitations in the paper relate to these? if at all...