I don't really see the relevance. As the article itself says, you can just add more dimensions and separating the clusters becomes trivial. And empirically unde

I don't really see the relevance. As the article itself says, you can just add more dimensions and separating the clusters becomes trivial. And empirically underfitting and local optima do not seem to be big issues for large neural networks.