BEGIN:VCALENDAR
METHOD:PUBLISH
VERSION:2.0
PRODID:-//Cornell U. Department of Computer Science//Brown Bag Seminar//EN
BEGIN:VEVENT
SUMMARY:Brown bag: Ziv Goldfeld
DESCRIPTION:Title: Smoothing probability distributions for high
dimensional learning and inference\nSpeaker: Ziv Goldfeld\nAbstract: The
talk will explore the benefits of smoothing probability distribution (by
convolving them with a chosen kernel) for high dimensional learning and
inference. Machine learning tasks often involve optimization of
evaluation of a certain functional of the underlying data distribution\,
e.g.\, loss function\, information measure\, statistical distance\, etc.
In practice\, we rarely have access to the actual distribution and only
get data from it. This necessitates estimating the distribution or the
functional of interest from samples. A central issue is that such
estimators suffer from the curse of dimensionality\, i.e.\, their sample
complexity grows exponentially fast with dimension. This makes it
impossible to obtain meaningful accuracy guarantees\, considering the
dimensionality of real-world data. As we shall see\, smoothing
alleviates the curse of dimensionally while preserving the capability to
perform inference by leveling out local irregularities in the considered
distributions. This enables constructing estimators with scalable (in
dimension) sample complexity guarantees and opens the door for various
applications. The talk will cover two such applications: measuring
information flows in deep neural network classifiers and implicit
generative modeling via minimum distance estimation. We will discuss the
original challenges\, how smoothing helps to deal with them\, remaining
gaps\, and ongoing/future research trajectories.
LOCATION:Zoom
UID:2020-12-01
STATUS:TENTATIVE
DTSTART:20201201T182500Z
DTEND:20201201T194000Z
LAST-MODIFIED:20201130T164344Z
ORGANIZER;CN=Jonathan Shi:http://www.cs.cornell.edu/~jshi/brownbag/
DTSTAMP:20240222T045640Z
END:VEVENT
END:VCALENDAR