A Positive Outlook on Negative Curvature

August 15, 2018, 10:15 AM - 10:45 AM

Location:

Iacocca Hall

Lehigh University

Bethlehem PA

Click here for map.

Daniel Robinson, Johns Hopkins University

The recent surge in interest in nonconvex models (e.g., in deep learning, subspace clustering, and dictionary learning) emphasizes a need for a fresh look at nonconvex optimization algorithms with provable convergence guarantees. A major factor in the design of such methods is the manner in which negative curvature is handled. In this talk, I present recent work that supports the following claims: (i) Commonly employed strategies for using negative curvature directions usually hurt algorithm performance; (ii) A new strategy based on upper-bounding models allows directions of negative curvature to be used while improving performance; and (iii) This strategy of using upper-bounding models is readily adapted for stochastic optimization, thus making it an attractive approach for large-scale "big data" problems. The talk also touches on worst-case complexity bounds and the pitfalls of attempting to associate such bounds with practical performance.