[Colloquium] Talks at TTIC

Dawn Ellis dellis at ttic.edu
Tue Mar 26 09:41:26 CDT 2013


When:  Monday, April 1st at 11:00am

Where: TTIC:  6045 S. Kenwood Ave. 5th Floor, Room 530

Speaker: Laurens van der Maaten
Affiliation: Delft University of Technology

Title: On Constructing and Corrupting Features

Abstract:
In this talk, I will give an overview of my work in machine learning.
Specifically, I will present my work on (1) constructing low-dimensional
feature representations that are suitable for visualization and (2)
corrupting feature representations in order to regularize learning models.
In addition, I will give a very brief introduction to my work on structured
prediction.

In the first half of the talk, I will present a new technique for learning
low-dimensional features that can be visualized in scatter plots, called
t-Distributed Stochastic Neighbor Embedding (t-SNE). I will show that t-SNE
is much better than alternative techniques in revealing the underlying
structure of the data. In addition, I will present an approach to scale up
t-SNE to big data and a variant of t-SNE that can be used to visualize
non-metric similarities.

In the second half of the talk, I will present a novel way of regularizing
classifiers and regressors, called marginalized corrupted features (MCF).
The key idea of MCF is to define a distribution over feature corruptions
that are believed to be label-invariant, and to train the model on
infinitely many samples from that distribution. I will show that MCF may
improve the performance of learners, e.g., in document classification, in
part-of-speech tagging under domain shift, and in the classification of
test data in which variables are missing at random.

Host: Gregory Shakhnarovich, greg at ttic.edu

-- 
*Dawn Ellis*
Administrative Assistant
773-834-1757
dellis at ttic.edu

TTIC
6045 S. Kenwood Ave.
Chicago, IL. 60637
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20130326/08f845d9/attachment.htm 


More information about the Colloquium mailing list