[AI Seminar] AI Seminar sponsored by Apple -- Veeranjaneyulu Sadhanala -- Dec 12

[Adams Wei Yu]

Dear faculty and students,

We look forward to seeing you next Tuesday, Dec 12, at noon in NSH 3305 for
AI Seminar sponsored by Apple. To learn more about the seminar series,
please visit the AI Seminar webpage <http://www.cs.cmu.edu/~aiseminar/>.

On Tuesday, Veeranjaneyulu Sadhanala <http://www.cs.cmu.edu/~vsadhana/>
will give the following talk:

Title: Escaping saddle points in neural network training and other
non-convex optimization problems

Abstract:

In non-convex optimization problems, first-order based methods can get
stuck at saddle points which are not even local minima. The generalization
error at saddle points is typically large and hence it is important to move
away from them. We discuss recently developed algorithms to escape saddle
points. In particular, we discuss gradient descent perturbed with additive
isotropic noise and Newton method with cubic regularization. They converge
to \epsilon-second order stationary points (informally, local minima) in
O(polylog(d) / \epsilon^2) time and O(1/ \epsilon^1.5) iterations
respectively under some conditions on the structure of the objective
function.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/ai-seminar-announce/attachments/20171210/fa91fcc6/attachment.html>