[Colloquium] Reminder: Mikhail Belkin Speaking Today @ TTI-C @ 3pm

[Katherine Cumming]

Guest Speaker
 
Speaker:  Mikhail Belkin, University of Chicago
Speaker's homepage:  http://www.cs.uchicago.edu/people/misha
 
Time:  Tuesday, March 29th @ 3:00pm
Location:  TTI-C Conference Room
Refreshments provided
 
Title:  Geometric Inference with Laplace Operator -- a New Framework for
Machine Learning
 
Abstract:  
I will discuss our work on Laplacians associated to point clouds and
will propose a systematic framework for using geometry of data in
machine learning problems. This approach produces a family of algorithms
for learning from labeled and unlabeled data.

In particular, I will discuss the following:

1. Data representation/visualization. In the last few years there has
been a considerable amount of interest in manifold learning algorithms.
I will describe our Laplacian Eigenmaps algorithm, which was the first
manifold learning algorithm with theoretical guarantees for arbitrary
manifolds. Applications, including machine vision, will be considered.

2. Semi-supervised learning. Kernel methods including Support Vector
Machines, are widely considered to be among best-performing algorithms
for supervised classification problems. However for many real-world
tasks, such as text mining, it is important to make use of unlabeled
data as labels often have to be expensively produced by hand. I will
describe a computationally efficient and theoretically satisfactory
method of incorporating unlabeled data into kernel methods and present
experimental comparisons showing strong performance on a variety of
datasets including text classification and optical character
recognition.

3. Clustering. I will discuss a geometric interpretation of spectral
clustering, which is a popular technique for partitioning data, and
present first statistical consistency results along with a new
algorithm.

4. If time permits, I will also mention regularizations on graphs and
some intriguing connections to learning Boolean circuits using Fourier
analysis. 
 
 
 
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20050329/ea373dea/attachment.htm