to write.

A nice thing to add is that if we think to the training of a neural network as the procedure of choosing among the various dynamical systems solving a task, a good one, we should aim to converge to structurally stable ones. In this way even if we converge to a local minimum, changes to the weights should not alter significantly the performance.

another point of view could be that maybe naturally this is the choice of an optimiser, since it would mean convergin to a relatively flat area of the loss function (read if there are works about this)