[Colloquium] 6/25 TTIC Colloquium: Greg Durrett, UT Austin

[Mary Marre via Colloquium]

 *When:*      Monday, June 25th at 10:30 am

*Where:    * TTIC, 6045 S. Kenwood Avenue, 5th Floor, Room 526

*Who:        *Greg Durrett, UT Austin


*Title:*        Hand-holding Learning of Latent Structure for Natural
Language Processing

*Abstract:* Neural network models are extremely powerful, but can typically
fit a training set in many ways, including ways that don't generalize; this
is the classic overfitting problems. While techniques like regularization
and dropout can help, these still don't explicitly encourage networks to
learn the right latent representations for the problem at hand.  In this
talk, I will discuss some recent work my group has done on "hand-holding"
neural network models for NLP: incorporating auxiliary supervision into
neural networks or constraining their latent representations to improve
generalization. I will primarily focus on two NLP problems involving
complex reasoning: document-based reading comprehension and GRE-style math
word problem solving. For both problems, our models must learn to perform
multiple steps of reasoning based on input text which can be both ambiguous
and display high amounts of lexical variation.  We incorporate auxiliary
supervision derived either from human annotation or a heuristic solver and
show that this can improve the performance of a learned model.  Finally, I
will briefly discuss some current work on choosing the right distribution
for latent variables in deep generative models, which imposes a similar
kind of prior knowledge about the task's structure.


Host:  Kevin Gimpel <kgimpel at ttic.edu>



Mary C. Marre
Administrative Assistant
*Toyota Technological Institute*
*6045 S. Kenwood Avenue*
*Room 504*
*Chicago, IL  60637*
*p:(773) 834-1757*
*f: (773) 357-6970*
*mmarre at ttic.edu <mmarre at ttic.edu>*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20180618/12a68e22/attachment.html>