All events

16 June 2017
Paris

« Subtle but not malicious? The (high) computational cost of non-smoothness in learning from big data »

The CREST (center of research in economy and statistics) organize a seminar titled « Subtle but not malicious? The (high) computational cost of non-smoothness in learning from big data » on statistics.

Speaker

Mikhail BELKIN, Ohio State University

organizer

Alexandre TSYBAKOV (Statistical Laboratory CREST)

Cristina BUTUCEA (Statistical Laboratory CREST)

Presentation from the speaker

« Subtle but not malicious? The (high) computational cost of non-smoothness in learning from big data »

What can we learn from big data? First, more data allows us to more precisely estimate probabilities of uncertain outcomes. Second, data provides better coverage to approximate functions more precisely. I will argue that the second is key to understanding the recent success of large scale machine learning. A useful way of thinking about this issue is that it is necessary to use many more components for principal component regression/classification, perhaps almost as many as data points. It turns out that there are fundamental computational barriers preventing some of the standard techniques (e.g., the kernel methods) from utilizing sufficiently many principal components on large datasets. These computational limitations result in over-regularization and failure to benefit from big data. I will discuss the nature of these barriers and how they can be overcome. In particular, I will show a simple kernel algorithm (EigenPro) demonstrating significant and consistent improvements over the state of the art on large datasets.

Based on joint work with Siyuan Ma.

INFORMATIONS

Friday June 16, 2017

11:00-12:15

adress

ENSAE

3 avenue Pierre Larousse

92240 Malakoff

transportation

Tramway T3 : “Porte de Vanves”

or Métro station 13 : “Porte de Vanves” or “Malakoff Plateau de Vanves”)

Life research

Welcome to the New CESSDA director

The Max Planck Institute for Demographic Research is seeking a Student Assistant for the Labor Demography project of the Laboratory of...

Opportunities

1 September 2017

Student Assistant at The Max Planck Institute for Demographic Research

The Max Planck Institute for Demographic Research is seeking a Student Assistant for the Labor Demography project of the Laboratory of Population Health.