Why not just dump the neural networks and deep learning? [closed]

manpreet Best Answer 3 years ago

Fundamental problem with deep learning and neural networks in general.

1. The solutions that fit training data are infinite. We don't have precise mathematical equation that is satisfied by only a single one and that we can say generalizes best. Simply speaking we don't know which generalizes best.

2. Optimizing weights is not a convex problem, so we never know we end up with a global or a local minimum.

So why not just dump the neural networks and instead search for a better ML model? Something that we understand, and something that is consistent with a set of mathematical equations? Linear and SVM do not have this mathematical drawbacks and are fully consistent with a a set of mathematical equations. Why not just think on same lines (need not be linear though) and come up with a new ML model better than Linear and SVM and neural networks and deep learning?