.png){kind=icon}
.png){kind=icon}
Return
Escaping from saddle points with perturbed gradient estimation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2026.131549.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Proposes a novel zeroth-order algorithm, 2-SPSA, for non-convex optimization. • Utilizes a structured sampling scheme to implicitly leverage Hessian information for escaping saddle points. • Demonstrates superior computational efficiency over existing methods in empirical studies.
Keywords:
zeroth-order optimization
saddle point escape
structured sampling
Hessian information
gradient estimation
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W