Date & Time:
February 25, 2019 10:30 am – 11:30 am
Location:
Crerar 390, 5730 S. Ellis Ave., Chicago, IL,
02/25/2019 10:30 AM 02/25/2019 11:30 AM America/Chicago Jason Lee (USC) – On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization Crerar 390, 5730 S. Ellis Ave., Chicago, IL,

On the Foundations of Deep Learning: SGD, Overparametrization, and Generalization

We provide new results on the effectiveness of SGD and overparametrization in deep learning.

a) SGD: We show that SGD converges to stationary points for general nonsmooth, nonconvex functions, and that stochastic subgradients can be efficiently computed via Automatic Differentiation. For smooth functions, we show that gradient descent, coordinate descent, ADMM, and many other algorithms, avoid saddle points and converge to local minimizers. For a large family of problems including matrix completion and shallow ReLU networks, this guarantees that gradient descent converges to a global minimum.

b) Overparametrization: We show that gradient descent finds global minimizers of the training loss of overparametrized deep networks in polynomial time. 

c) Generalization: For general neural networks, we establish a margin-based theory. The minimizer of the cross-entropy loss with weak regularization is a max-margin predictor, and enjoys stronger generalization guarantees as the amount of overparametrization increases. 

d) Algorithmic and Implicit Regularization: We analyze the implicit regularization effects of various optimization algorithms on overparametrized networks. In particular we prove that for least squares with mirror descent, the algorithm converges to the closest solution in terms of the bregman divergence. For linearly separable classification problems, we prove that the steepest descent with respect to a norm solves SVM with respect to the same norm. For over-parametrized non-convex problems such as matrix sensing or neural net with quadratic activation, we prove that gradient descent converges to the minimum nuclear norm solution, which allows for both meaningful optimization and generalization guarantees.

Host: Rebecca Willett

Jason Lee

Assistant Professor, University of Southern California

Jason Lee is an assistant professor of Data Sciences and Operations in Marshall School of Business, Computer Science Department (courtesy), and a member of the Machine Learning Center. Previously, he was a postdoc in the Computer Science Department at UC Berkeley working with Michael Jordan, and also collaborated with Ben Recht. Prior to that, he was a PhD student in Computational Mathematical Engineering at Stanford University advised by Trevor Hastie and Jonathan Taylor.  He received a BS in Mathematics from Duke University advised by Mauro Maggioni.  His research interests are broadly in Statistics (Statistical Learning and High-dimensional Statistics), Machine Learning, and Optimization. 

Related News & Events

Video

“Machine Learning Foundations Accelerate Innovation and Promote Trustworthiness” by Rebecca Willett

Jan 26, 2024
Video

Nightshade: Data Poisoning to Fight Generative AI with Ben Zhao

Jan 23, 2024
No Name

In The News: U.N. Officials Urge Regulation of Artificial Intelligence

"Security Council members said they feared that a new technology might prove a major threat to world peace."
Jul 27, 2023
No Name

UChicago Computer Scientists Bring in Generative Neural Networks to Stop Real-Time Video From Lagging

Jun 29, 2023
No Name

Computer Science Displays Catch Attention at MSI’s Annual Robot Block Party

Apr 07, 2023
No Name

UChicago, Stanford Researchers Explore How Robots and Computers Can Help Strangers Have Meaningful In-Person Conversations

Mar 29, 2023
Students posing at competition
No Name

UChicago Undergrad Team Places Second Overall In Regionals For World’s Largest Programming Competition

Mar 17, 2023
No Name

Postdoc Alum John Paparrizos Named ICDE Rising Star

Mar 15, 2023
No Name

New EAGER Grant to Asst. Prof. Eric Jonas Will Explore ML for Quantum Spectrometry

Mar 03, 2023
No Name

Assistant Professor Chenhao Tan Receives Sloan Research Fellowship

Feb 15, 2023
No Name

UChicago Scientists Develop New Tool to Protect Artists from AI Mimicry

Feb 13, 2023
No Name

Professors Rebecca Willett and Ben Zhao Discuss the Future of AI on Public Radio

Jan 26, 2023
arrow-down-largearrow-left-largearrow-right-large-greyarrow-right-large-yellowarrow-right-largearrow-right-smallbutton-arrowclosedocumentfacebookfacet-arrow-down-whitefacet-arrow-downPage 1CheckedCheckedicon-apple-t5backgroundLayer 1icon-google-t5icon-office365-t5icon-outlook-t5backgroundLayer 1icon-outlookcom-t5backgroundLayer 1icon-yahoo-t5backgroundLayer 1internal-yellowinternalintranetlinkedinlinkoutpauseplaypresentationsearch-bluesearchshareslider-arrow-nextslider-arrow-prevtwittervideoyoutube