Statistics and Probability Seminars

Statistics Department Seminar: Modern surprises in classical machine learning

Topic: 
Modern surprises in classical machine learning
Abstract / Description: 

Seemingly counter-intuitive phenomena in deep neural networks and kernel methods have prompted a recent re-investigation of classical machine learning methods, like linear models. Of particular focus is sufficiently high-dimensional setups in which interpolation of training data is possible. In this talk, we will first review recent works showing that zero regularization, or fitting of noise, need not be harmful in regression tasks. Then, we will use this insight to uncover two new surprises for high-dimensional linear classification:

  • least-2-norm interpolation can classify consistently even when the corresponding regression task fails, and
  • the support-vector-machine will exactly interpolate discrete labels, i.e., all training points become support vectors, in sufficiently high-dimensional models.

This is joint work with Misha Belkin, Daniel Hsu, Adhyyan Narang, Anant Sahai, Vignesh Subramanian, and Ji Xu.

Date and Time: 
Tuesday, February 23, 2021 - 4:30pm

Statistics Department Seminar: Distribution-free, risk-controlling prediction sets

Topic: 
Distribution-free, risk-controlling prediction sets
Abstract / Description: 

To enable valid statistical inference in prediction tasks, we show how to generate set-valued predictions with black-box models that control various notions of statistical error. Our approach guarantees that the expected loss on future test points falls below a user-specified level, for any predictive model and underlying distribution. Building on conformal prediction, we use a holdout set to calibrate the size of the prediction sets, generalizing the approach to control error notions such as the false rejection rate. We demonstrate our procedure in four large-scale problems: (1) multi-label classification, where each observation has multiple associated labels; (2) classification problems where the labels have a hierarchical structure; (3) image segmentation, where we wish to predict a set of pixels containing an object of interest; and (4) protein structure prediction.

Date and Time: 
Tuesday, February 16, 2021 - 4:30pm

Probability Seminar of the Americas presents "Structure theorems for information in streamed data"

Topic: 
Structure theorems for information in streamed data
Abstract / Description: 

A basic question is to understand the space of real valued functions on the space of unparametrized path segments. I will explain that there are atomic ways to uniquely factor the space of "polynomial" functions on streams into two parts, a potentially expensive to compute information tensor, and a space of quick to compute polynomial functions on this informative tensor. The approach is atomic in the sense that the information in an atom from the tensor can be computed from the data without having to compute the full information. This makes the result of great potential value for situations where dimension is critical. The proofs are pure algebra. We explain that hall integrals, and hall areas are examples of uniquely informative tensors.

This work is primarily that of Cris Salvi with support from Joscha Diehl,
Terry Lyons, Rosa Preiß, Jeremy Reizenstein.


 

The seminar will take place on Wednesdays and will usually consist of one talk at 11am Pacific Standard Time (2pm Eastern Standard Time). We aim for two or more talks per month. The first speaker will be Terry Lyons of Oxford University

The purpose of the seminar is to bring together researchers in probability theory, catering primarily to the North and South American Time zones, and to allow time to gather socially at a "virtual cafe" (bring your own drink) after the seminar. (Here people can walk around with a virtual avatar and interact with groups-including shared white boards, etc.)

The seminar will be using the Zoom platform, and afterwards a link will be given for a virtual social gathering taking place on a different platform. Talks will be recorded and available on YouTube (with permission).

If you would like to get future announcements you can either follow the seminar website and calendar or join the seminar mailing group.

 

Date and Time: 
Wednesday, February 3, 2021 - 11:00am to Thursday, February 4, 2021 - 10:55am

Statistics Department Seminar welcomes Subhabrata Sen

Topic: 
TBA
Abstract / Description: 

The Statistics Seminars for Winter Quarter will be held online via Zoom at 4:30pm on Tuesdays. Please note that these events will be locked by the Host after ten minutes, so latecomers will not be able to join the Meeting. Subscribe to Statistics distribution list to receive Meeting IDs and passwords via email.

Date and Time: 
Tuesday, March 16, 2021 - 4:30pm

Statistics Department Seminar welcomes Jonathan Niles-Weed

Topic: 
TBA
Abstract / Description: 

The Statistics Seminars for Winter Quarter will be held online via Zoom at 4:30pm on Tuesdays. Please note that these events will be locked by the Host after ten minutes, so latecomers will not be able to join the Meeting. Subscribe to Statistics distribution list to receive Meeting IDs and passwords via email.

Date and Time: 
Tuesday, March 9, 2021 - 4:30pm

Statistics Department Seminar welcomes Yanjun Han

Topic: 
Learning to bid in repeated first-price auctions
Abstract / Description: 

First-price auctions have very recently swept the online advertising industry, replacing second-price auctions as the predominant auction mechanism on many platforms. This shift has brought forth important challenges for a bidder: how should one bid in a first-price auction where it is no longer optimal to bid one's private value truthfully and hard to know the others' bidding behaviors? To answer this question, we study online learning in repeated first-price auctions, and consider various scenarios involving different assumptions on the characteristics of the other bidders' bids, of the bidder's private valuation, of the feedback structure of the auction, and of the reference policies with which our bidder competes. For all of them, we characterize the essentially optimal performance and identify computationally efficient algorithms achieving it. Experimentation on first-price auction datasets from Verizon Media demonstrates the promise of our schemes relative to existing bidding algorithms.

This is based on joint work with Aaron Flores, Erik Ordentlich, Tsachy Weissman, and Zhengyuan Zhou.


 

 

The Statistics Seminars for Winter Quarter will be held online via Zoom at 4:30pm on Tuesdays. Please note that these events will be locked by the Host after ten minutes, so latecomers will not be able to join the Meeting. Subscribe to Statistics distribution list to receive Meeting IDs and passwords via email.

Date and Time: 
Tuesday, February 9, 2021 - 4:30pm

Statistics Department Seminar welcomes Daniel Erdmann-Pham

Topic: 
TBA
Abstract / Description: 

The Statistics Seminars for Winter Quarter will be held online via Zoom at 4:30pm on Tuesdays. Please note that these events will be locked by the Host after ten minutes, so latecomers will not be able to join the Meeting. Subscribe to Statistics distribution list to receive Meeting IDs and passwords via email.

Date and Time: 
Tuesday, February 2, 2021 - 4:30pm

Statistics Department Seminar presents "Hypothesis testing for large-scale data: Enhancing reliability and efficiency"

Topic: 
Hypothesis testing for large-scale data: Enhancing reliability and efficiency
Abstract / Description: 

In scientific research that involves large-scale data, researchers often start with questions regarding the global properties of a large set of measurements. For instance, are a group of related genes in the same functional pathway jointly associated with a trait of interest? Such questions can be formulated as hypothesis testing problems that globally examine a large number of parameters in a high-dimensional joint distribution. Examples include hypothesis testing on mean vectors, covariance matrices and regression coefficients. To extract informative scientific knowledge from abundant data, reliability and efficiency are among the major concerns in statistical inference.

In this talk, I will address particular reliability and efficiency issues arising from jointly testing a large number of parameters. First, I will discuss how reliable the popular likelihood ratio tests (LRTs) are in terms of the type I error control for high-dimensional data. I will provide theoretical insights into the reliability of the LRTs in a variety of problems, which are based on phase transition results of the foundational Wilk's theorem. Next, to improve efficiency of the existing testing procedures under high-dimensional settings, I will introduce a new adaptive testing framework that can maintain high statistical power against a wide range of alternative hypotheses. The proposed framework is based on a family of U-statistics that are constructed to capture the information in different directions in high-dimensional spaces. For a broad class of problems, we establish high-dimensional asymptotic theory for the U-statistics and develop adaptive testing procedures that are statistically powerful in a wide variety of scenarios.

 


 

The Statistics Seminars for Winter Quarter will be held online via Zoom at 4:30pm on Tuesdays. Please note that these events will be locked by the Host after ten minutes, so latecomers will not be able to join the Meeting. Subscribe to Statistics distribution list to receive Meeting IDs and passwords via email.

Date and Time: 
Tuesday, January 26, 2021 - 4:30pm

Statistics Department Seminar welcomes Somabha Mukherjee

Topic: 
TBA
Abstract / Description: 

The Statistics Seminars for Winter Quarter will be held online via Zoom at 4:30pm on Tuesdays. Please note that these events will be locked by the Host after ten minutes, so latecomers will not be able to join the Meeting. Subscribe to Statistics distribution list to receive Meeting IDs and passwords via email.

Date and Time: 
Tuesday, January 19, 2021 - 4:30pm

Pages

Subscribe to RSS - Statistics and Probability Seminars