4 ms·

Well, in general, no, there is no perturbation method large (or good) enough to get out of saddle points, including via stochastic gradient descent. It might wo

by Cacti 9y ago

Well, in general, no, there is no perturbation method large (or good) enough to get out of saddle points, including via stochastic gradient descent. It might work in a particularly specific problem, but not in general.

Eridrus 9y ago
Fine, but I mostly care about ML applications, where I'm wondering if this is expected to help at all.
- Cacti 9y ago
  ML is far too broad a category. Like I said, it will depend on the problem/function.
  - robrenaud 9y ago
    How does what you say agree/disagree with the authors paper? The authors bound derivatives and 2nd derivatives, which excludes particularly nasty functions?