Fwd: [Fwd: Special Machine Learning & Algorithms Seminar, Ravi Kannan - February 24, 2015]

[Barnabas Poczos]

FYI


---------- Forwarded message ----------
From: Nina Balcan <ninamf at cs.cmu.edu>

Hey guys.

Don't miss this! It is supposed to be great.
Barnabas and Aarti, please forward it to your students.

Nina
---------------------------- Original Message ----------------------------
Subject: [Fwd: Special Machine Learning & Algorithms Seminar,

Hi everyone.

Please join us for a Special Machine Learning & Algorithms Seminar by
Ravi Kannan, at 2:00 PM tomorrow. Please see full details below.

Best.
Nina Balcan

---------------------------- Original Message ----------------------------

Please join us for a Special Machine Learning & Algorithms Seminar!

Tuesday, February 24, 2015
2:00 pm
NSH  3305


Ravi Kannan<http://research.microsoft.com/en-us/people/kannan/>, Principal
Researcher in the Algorithms - Microsoft Research
Host: Nina Balcan


Title:   Topic Modeling: A Provable Algorithm

Abstract:

Topic Modeling is widely used. The general model inference problem is
hard. Known provable algorithms need some strong assumptions on the model.

After a from-first-principles introduction, the talk will formulate a
model with two natural assumptions: (i) each document has a dominant topic
and (ii) each topic has dominant words. We provably solve the model
fitting problem. The algorithm is based on three natural components: a
thresholding step, Singular Value Decomposition and clustering. The proof
crucially draws upon recent results on the convergence of another widely
used tool: Lloyd's k-means algorithm. We test both the performance of the
algorithm and the validity of the assumptions on real document corpora
with success. The simple device of thresholding has other uses - we will
see an exponential advantage for certain \planted" problems in terms of
the signal-to-noise ratio.

Joint with T. Bansal and C. Bhattacharyya, Indian Institute of Science.

Bio

Ravindran (Ravi) Kannan is Principal Researcher in the Algorithms Research
Group at Microsoft Research Bangalore. Previously he was a professor at
CMU, MIT, and Yale, where he was the William Lanman Professor of Computer
Science. His research areas span Algorithms, Optimization and Probability.

He is widely known for introducing several groundbreaking techniques in
theoretical computer science, notably in the algorithmic geometry of
numbers, sampling and volume computation in high dimension, and
algorithmic linear algebra. He received the Knuth Prize in 2011, and the
Fulkerson Prize in 1992. He is a distinguished alumnus of IIT Bombay.