Category Archives: Winter 2018

Walid Krichene; Continuous-time dynamics for convex optimization

Feb 20th, 2018, 12:00pm
CSE 403
Walid Krichene, Google

Abstract: Many optimization algorithms can be viewed as a discretization of a continuous-time process (described using an ordinary differential equation, or a stochastic differential equation in the presence of noise). The continuous-time point of view can be useful for simplifying the analysis, drawing connections with physics, and streamlining the design of new algorithms and heuristics. We will review results from continuous-time optimization, from the simple case of gradient flow, to more recent results on accelerated methods. In particular, we give simple interpretations of acceleration, and show how these interpretations motivate heuristics (restarting and adaptive averaging) which, empirically, can significantly improve the speed of convergence. We will then focus on the stochastic case, and study the interaction between acceleration and noise, and their effect on the convergence rates. We will conclude with a brief review of how the same tools can be applied in other problems, such as approximate sampling and non-convex optimization.

Bio: Walid Krichene is at Google Research, where he works on large-scale optimization and recommendation. He received his Ph.D. in EECS in 2016 from UC Berkeley, where he was advised by Alex Bayen and Peter Bartlett, a M.A. in Mathematics from U.C. Berkeley, and a M.S. in Engineering and Applied Math from the Ecole des Mines ParisTech. He received the Leon Chua Award and two outstanding GSI awards from U.C. Berkeley. His research interests include convex optimization, stochastic approximation, recommender systems, and online learning.

Jakub Konečný; Federated Learning: Privacy-Preserving Collaborative Machine Learning without Centralized Training Data

Jan 30th, 2018, 12:00pm
CSE 403
Jakub KonečnýGoogle

Abstract: Federated Learning is a machine learning setting where the goal is to
train a high quality centralized model while training data remains
distributed over a large number of clients each with unreliable and
relatively slow network connections. We consider learning algorithms
for this setting where on each round, each client independently
computes an update to the current model based on its local data, and
communicates this update to a central server, where the client-side
updates are aggregated to compute a new global model.

In this talk, I will introduce the underlying algorithms, and present
several ideas for improving the overall system in terms of
communication efficiency, security, and differential privacy.

Bio: Jakub Konečný is a research scientist at Google working on Federated
Learning, an effort to decentralize machine learning. Prior to joining
Google, Jakub completed his PhD at University of Edinburgh focusing on
optimization algorithms for machine learning.

Zeyuan Allen-Zhu; How to Swing By Saddle Points: Faster Non-Convex Optimization Than SGD

Feb 13th, 2018, 4pm
LOW 105
Zeyuan Allen-Zhu, Microsoft Research

“The diverse world of deep learning research has given rise to thousands of architectures for neural networks. However, to this date, the underlying training algorithms for neural networks are still stochastic gradient descent (SGD) and its heuristic variants.

In this talk, we present a new stochastic algorithm to train any smooth neural network to ε-approximate local minima, using O(ε^{-3.25}) backpropagations. The best provable result was O(ε^{-4}) by SGD before this work.

More broadly, it finds ε-approximate local minima of any smooth nonconvex function using only O(ε^{-3.25}) stochastic gradient computations.”