<video controls width="1280" height="720" poster="http://streamod.cs.brown.edu:8801/K/4/hires.jpg" > <source type="video/mp4" src="http://streamod.cs.brown.edu:8801/K/4/hires.mp4" /> <source type="video/ogg" src="http://streamod.cs.brown.edu:8801/K/4/hires.ogv" /> <applet code="com.fluendo.player.Cortado.class" archive="/cortado/cortado.jar" width="1280" height="720"><param name="url" value="http://streamod.cs.brown.edu:8801/K/4/hires.ogv"/></applet></video>
Star Convexity and Robust Deep Learning: Two Perspectives from
Stochastic Optimization
I will discuss two recent research projects aiming at a better
understanding of stochastic optimization algorithms. The first project
tackles the challenge of "star-convex optimization", developing a new
randomized algorithm that for the first time allows efficient
optimization of a natural class of non-convex functions arising in
machine learning contexts. The second project is ongoing work
attempting to explain the mystery "why does deep learning generalize
so well?" We reveal a previously unknown side effect of stochastic
gradient descent that, under very broad conditions, nudges deep
learning models towards simpler hypotheses that generalize better.
Host: Professor Ugur Cetintemel