How to Escape Saddle Points Efficiently? by Praneeth Netrapalli

Explore the intricacies of escaping saddle points efficiently in this 51-minute conference talk by Praneeth Netrapalli at the International Centre for Theoretical Sciences. Delve into non-convex optimization techniques, examining two major observations and the current state of the art. Gain insights into perturbed gradient descent, analyzing its application in two-dimensional, three-dimensional, and general quadratic cases. Understand the key ingredients and proof ideas behind efficient saddle point escape methods. Conclude with a discussion on open questions in the field, providing a comprehensive overview of this crucial topic in algorithms and optimization.

Intro
Non-convex optimization
Two major observations
State of the art
Summary of results
Setting
Perturbed gradient descent
Key question
Two dimensional quadratic case
Three dimensional quadratic case
General case
Two key ingredients of the proof
Proof idea
Putting everything together
Open questions