[Crib-list] TODAY... SPEAKER: Stefanie Jegelka (MIT) | CRIBB Seminar | Friday, Nov. 2, 2018 | TIME: 1:00 PM | Room 32-155 (Stata)

[Shirley Entzminger]

 	T O D A Y . . .



 		COMPUTATIONAL RESEARCH in BOSTON and BEYOND Seminar



DATE:		Friday, November 2, 2018

TIME:		1:00 PM - 2:00 PM

LOCATION:	Building 32, Room 155  (STATA)
 		  (32 Vassar Street, Cambrdidge, MA)

 		    Pizza and beverages provided outside
          	    Room 155 at 12:45 PM


TITLE:	 Aspects of Robustness and Representation in Machine Learning


SPEAKER:	 Stefanie Jegelka  (MIT)


ABSTRACT:

Reliability of machine learning methods includes many facets. One aspect are 
robust, stable algorithms. Another one is a better theoretical understanding of 
properties of currently popular models. In this talk, I will show recent work 
on both these directions.

First, we address robustness for black-box optimization with Gaussian Processes 
(GPs). GP-based methods have become popular tools for sequentially optimizing 
an unknown function that is expensive to evaluate, with applications in 
robotics, hyperparameter tuning, recommender systems and environmental 
monitoring. In such applications, robust, stable solutions are of interest for 
several reasons: the underlying functions during optimization and 
implementation stages are different, or one seeks an entire region of good 
inputs rather than only a single point. We formalize this by allowing the query 
point to be adversarially perturbed, and require the function value to remain 
as high as possible even after this perturbation. Standard GP optimization 
approaches can fail in this setting. We provide a new, confidence-bound based 
algorithm, and establish lower and upper bounds on the required number of 
samples to find a near-optimal point.

Second, we explore the representational power of ResNet, a popular recent 
neural network architecture that augments the network with a parallel identity 
mapping. While classical results address wide, shallow networks, we ask how 
narrow a deep ResNet can be to still allow universal approximation. Our results 
show that one hidden unit is enough, in sharp contrast to fully connected 
networks.


This talk is based on joint work with Ilija Bogunovic, Jonathan Scarlett, 
Volkan Cevher and Hongzhou Lin.

--------------------------------------------------------------------------

Massachusetts Institute of Technology
Cambridge, MA



For information about the Computational Research in Boston and Beyond Seminar 
(CRIBB), please visit...

                 http://math.mit.edu/crib/

===============

Shirley A. Entzminger
Administrative Assistant II
Department of Mathematics
Massachusetts Institute of Technology
77 Massachusetts Avenue
Building 2, Room 350A
Cambridge, MA 02139
PHONE:  (617) 253-4347
FAX:    (617) 253-4358
E-mail: daisymae at math.mit.edu
-------------- next part --------------
A non-text attachment was scrubbed...
Name: CRIBB Seminar - Friday, Nov 2-2018 - speaker,	Stefani Jegelka (MIT).pdf
Type: application/pdf
Size: 930460 bytes
Desc: 
Url : http://mailman.mit.edu/pipermail/crib-list/attachments/20181102/75c25683/attachment.pdf