While a good point that simple is relative, if you specifically look at deep neural networks we don't understand how training a non convex function converges wi

While a good point that simple is relative, if you specifically look at deep neural networks we don't understand how training a non convex function converges with gradient descent - the fundamental component to create a usable model. In practice, it often works, and there are a few intuitions why this works. But its naive at best to claim that this is simple. If it was we would understand it better :)