BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Deep (Inter-)Active Learning for NLP: Cure-all or Catastrophe? - Z
 achary Chase Lipton\, Carnegie Mellon University
DTSTART:20200124T150000Z
DTEND:20200124T160000Z
UID:TALK138658@talks.cam.ac.uk
CONTACT:Microsoft Research Cambridge Talks Admins
DESCRIPTION:While deep learning produces supervised models with unpreceden
 ted predictive performance on many tasks\, under typical training procedur
 es\, advantages over classical methods emerge only with large datasets. Th
 e extreme data-dependence of reinforcement learners may be even more probl
 ematic. Millions of experiences sampled from video-games come cheaply\, bu
 t human-interacting systems can’t afford to waste so much labor. In this
  talk\, I will discuss several efforts to increase the labor-efficiency of
  learning from human interactions. Specifically\, I will cover work on lea
 rning dialogue policies\, deep active learning for natural language proces
 sing\, learning from noisy and singly-labeled data\, and active learning w
 ith partial feedback. Finally\, time permitting\, I’ll discuss a new app
 roach for reducing the reliance of NLP models on spurious associations in 
 the data that relies on a new mechanism for interacting with annotators. 
LOCATION:Auditorium\, Microsoft Research Ltd\, 21 Station Road\, Cambridge
 \, CB1 2FB
END:VEVENT
END:VCALENDAR