Fanny Yang (Berkeley)

May 4, 2-3PM, 400 Cory.

Title and Abstract

Statistics meets optimization - computational guarantees for statistical learning algorithms
Modern technological advances have prompted massive scale data collection in many fields. This has led to an increasing need for scalable machine learning algorithms and statistical methods to draw conclusions about the world. Two principal challenges arise in this context: How can we collect data efficiently such that a reduced sample size is enough to draw conclusion with high confidence? How should we design the learning algorithm and how long should we run it? In this talk, I will address these questions for concrete problems in estimation and testing.

I will start by discussing the regularization effect of stopping gradient-type methods before convergence for non-parametric estimation. In particular, we show how both early stopping and penalty regularization can be explained by localized complexities. In the second part of the talk, I will elaborate on the advantages of efficient data collection and demonstrate how adaptive algorithms can reduce the number of required samples for simultaneous false discovery rate control and best-arm detection in multiple testing.

Bio

Fanny Yang is a graduate student in the PhD program in EECS, Berkeley, advised by Martin Wainwright. Her research interests focus on computational and statistical aspects of machine learning procedures aimed at obtaining good generalization. She has explored these issues in the context of non-convex optimization, non-parametric estimation and active learning. She received her B.A. degree in Electrical Engineering from Karlsruhe Institute of Technology and her M.Sc. degree from the Technical University Munich