BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Active visual category learning - Sudheendra Vijayanarasimhan
DTSTART:20110509T090000Z
DTEND:20110509T100000Z
UID:TALK31267@talks.cam.ac.uk
CONTACT:Microsoft Research Cambridge Talks Admins
DESCRIPTION:Visual recognition research develops algorithms and representa
 tions to autonomously recognize visual entities such as objects\, actions\
 , and attributes. The traditional protocol involves manually collecting tr
 aining image examples\, annotating them in specific ways\, and then learni
 ng models to explain the annotated examples. However\, this is a rather li
 mited way to transfer human knowledge to visual recognition systems\, part
 icularly considering the immense number of visual concepts that are to be 
 learned.\n\nI propose new forms of active learning that facilitate large-s
 cale transfer of human knowledge to visual recognition systems in a cost-e
 ffective way. The approach is cost-effective in the sense that the divisio
 n of labor between the machine learner and the human annotators respects a
 ny cues regarding which annotations would be easy (or hard) for either par
 ty to provide. The approach is large-scale in that it can deal with a larg
 e number of annotation types\, multiple human annotators\, and huge pools 
 of unlabeled data. In particular\, I consider three important aspects of t
 he problem: \n\n(1) cost-sensitive multi-level active learning\, where the
  expected informativeness of any candidate image annotation is weighed aga
 inst the predicted cost of obtaining it in order to choose the best annota
 tion at every iteration.\n(2) budgeted batch active learning\, a novel act
 ive learning setting that perfectly suits automatic learning from crowd-so
 urcing services where there are multiple annotators and each annotation ta
 sk may vary in difficulty.\n(3) sub-linear time active learning\, where on
 e needs to retrieve those points that are most informative to a classifier
  in time that is sub-linear in the number of unlabeled examples\, i.e.\, w
 ithout having to exhaustively scan the entire collection.\n\nUsing the pro
 posed solutions for each aspect\, I then demonstrate a complete end-to-end
  active learning system for scalable\, autonomous\, online learning of obj
 ect detectors. The approach provides state-of-the-art recognition and dete
 ction results\, while using minimal total manual effort. Overall\, my work
  enables recognition systems that continuously improve their knowledge of 
 the world by learning to ask the right questions of human supervisors.\n
LOCATION:Small lecture theatre\, Microsoft Research Ltd\, 7 J J Thomson Av
 enue (Off Madingley Road)\, Cambridge
END:VEVENT
END:VCALENDAR