Natural language research announcements

["Samdani, Rajhans" <rsamdan2 AT illinois.edu>]

Dear all,

Please see below the announcement for a talk by Hossein Mobahi at the digital 
signal processing seminar. The talk is very relevant to ML, AI, and Computer 
Vision people. Details:

TIME & PLACE: DSP (Digital Signal Processing) seminar, at 4:00-5:00
p.m. on Wednesdays April 25, Room: 4269 Beckman Institute.

TITLE : Seeing through the Blur

ABSTRACT: My talk will address the role of smoothing in nonconvex
optimization. When optimizing a nonconvex objective, a popular
heuristic is to smooth the objective function to hopefully make it
convex and easy to solve. Then while deforming the objective back to
the original, one follows the path of that minimizer back. This
heuristic often leads to a reasonable if not globally optimal
solution. The application I target is image alignment, but the
underlying optimization concept is quite generic and could be of
interest to AI and specially ML community.

The talk consists of two parts. I will first present some theoretical
results about the generic optimization setting; providing conditions
on functions for which Gaussian smoothing turns them convex, and also
present a closed form for their asymptotic minimizer (i.e. minimizer
after extremely smoothed). In the second part, I will focus on the
alignment application and discuss how to smooth the alignment
objective function. Smoothing requires a high dimensional integration
(due to Gaussian convolution) and thus is computationally expensive.
However, I will show that such integrals can be equivalently and
exactly computed by integration in a lower dimensional space via what
I call "blur kernels".

RESOURCES: The first (theoretical) part of the talk is not published
yet, but for the second, there is an associated paper for those
interested:
http://www.cs.illinois.edu/homes/hmobahi2/pubs/blur_cvpr12.pdf


BIO: Hossein Mobahi is a PhD student in CS department advised by Prof.
Yi Ma. He has done research on image segmentation, face recognition,
and 3d reconstruction. His current research focuses on the role of
smoothing in nonconvex optimization, specially those that arise in
computer vision.

Thanks!
Rajhans