Berkeley Statistics Student Seminar

The aim of this seminar is to give students a chance to give talks and inform their peers of the research that they are currently involved in. It also gives an opportunity to students to practice talks before conferences, job talks talks before their qualifying exams and so on. The seminar is generously funded by the Graduate Assembly and the Statistics department. For the Spring 2008 please send me or Richard (rhliang at stat) an email if you would like to give a talk. This seminar is being organized jointly by Bradley Luen, Chris Haulk, Richard Liang and myself. All the talks will be held in 1011 Evans from 4-5 p.m.

February 22	Chongsoon Bae	On Random Forests
February 29	Jing Lei	Some statistical issues in numerical weather forecasting
April 11	Charlotte Wickham	Modeling trajectories with an application in eye tracking .
April 18	Jian Ding	Cutoff of birth-and-death chains, and mixing-time of near critical models .
April 25	Irma Hernandez	Seasonality Volatility: Applications to Time Series and Point Processes
May 2	Nate Coehlo	TAOS and Some Relevant Statistical Problems

Abstracts

February 22

Speaker

Chongsoon Bae

Title

On Random Forests

Abstract

In recent years, many methods for regression and classification have been developed. The method of Random Forests is one of the most popular and it is well known that it provides good performance and is easy to calculate in many fields. But most of the results are based on empirical research and there were not sufficient analytic results.

In this presentation, I will start to explain CART, bagging. Then I will introduce the original Random Forests method (Leo Breiman's version) and other versions. If time permits, I hope to present related results for Random Forests algorithms.

February 29

Speaker

Jing Lei

Title

Some statistical issues in numerical weather forecasting

Abstract:

The statistical problem in numerical weather forecasting can be viewed as a prediction procedure in high-dimensional state-space model, which involves combining the background model predictions with noisy observations, a.k.a., data assimilation. In this talk I will introduce the background and briefly look into the state-of-art data assimilation method: ensemble Kalman Filter (EnKF). Two versions of EnKF are compared in the sense of robustness. Other issues such as dimension reduction and the use of particle filters will also be discussed.

April 11

Speaker

Charlotte Wickham

Title

Modeling trajectories with an application in eye tracking

Abstract:

April 18th

Speaker

Jian Ding

Title

Cutoff of birth-and-death chains, and mixing-time of near critical models

Abstract:

I will mainly talk about the following two problems, in a brief way. 1) Diaconis and Saloff-Coste (2006) have shown that for a family of continuous-time birth-and-death chains, started at an endpoint, the cutoff exists if and only if the product of the spectral gap and the mixing time tends to infinite, with convergence measured in separation distance. Recently, we have confirmed this equivalent condition for both continuous-time and lazy discrete-time birth and death chains, with convergence measured via total variation distance.

2, Recently, Levin, Luczak and Peres established results on mixing time of Glauber dynamics of Ising model on complete graph: the existence of cutoff and the order of mixing-time in different temperature regime, while $\beta$ is fixed. We study the case when $\beta$ is not fixed but tends to the critical value $1$.

Joint work with Eyal Lubetzky and Yuval Peres.

If time remains, I may also quickly mention another problem we have tried but have no results yet: the mixing time of the giant component of Erdos-Renyi graph $G_{n,p}$, in which $p=(1+\epsilon)/n$, when $\epsilon$ tends to $0$ and $\epsilon3 n$ goes to infinity. With Eyal Lubetzky, Asaf Nachmias and Yuval Peres.

April 25

Speaker

Irma Hernandez

Title

Seasonality and Volatility: Applications to Time Series and Point Processes

Abstract:

The modeling of seasonality and volatility is of interest in a variety of areas like: economics, risk analysis, ecology, transportation, finance… A review of parametric, non-parametric and hybrid methods for modeling seasonality and volatility is presented. Two applications are explored, one in Banking and the other in Forest Fires.

May 2

Speaker

Nate Coehlo

Title

TAOS and Some Relevant Statistical Problems

Abstract:

Modern Astronomy Surveys generate large quantities of data, so many statistical problems naturally arise. I will talk about one particular survey (TAOS), its scientific goals, data pipeline, current methodology, and areas where things can be improved. I will then discuss some relevant statistical methodology, including Multivariate Regression (Y Multivariate), Copulas, and Multivariate Volatility Modeling.