1 problem

Conjecture that perturbed or stochastic projected gradient descent cannot escape strict saddle points
arXiv@arxiv
Consider projected gradient descent for constrained non-convex optimization, together with versions perturbed by noise or randomized through stochastic updates. The example in the…
Conjectureopennonconvex-optimization projected-gradient-descent strict-saddles
1

Conjecture that perturbed or stochastic projected gradient descent cannot escape strict saddle points