Nightmare at Test Time: Robust Learning by Feature Deletion - PowerPoint PPT Presentation

1 / 11
About This Presentation
Title:

Nightmare at Test Time: Robust Learning by Feature Deletion

Description:

Nightmare at Test Time: Robust Learning by Feature Deletion Amir Globerson Computer Science and Articial Intelligence Laboratory, MIT, Cambridge, MA, USA – PowerPoint PPT presentation

Number of Views:98
Avg rating:3.0/5.0
Slides: 12
Provided by: geoc78
Category:

less

Transcript and Presenter's Notes

Title: Nightmare at Test Time: Robust Learning by Feature Deletion


1
Nightmare at Test Time Robust Learning by
Feature Deletion
  • Amir Globerson
  • Computer Science and Articial Intelligence
    Laboratory, MIT, Cambridge, MA, USA
  • Sam Roweis
  • Department of Computer Science, University of
    Toronto, Canada
  • All the equation images are clipped from this
    paper.

????????? (IBM???????/?????????1?)
2
???????
  • ????(?)????
  • Nightmare at Test Time
  • ? ???????????????
  • ??????Robust??
  • ??????Over weighting?????
  • ???????
  • ???????(????)??????????????
  • ??????????????????

3
???????????????
  • ????????????????????????
  • Spam filtering ????????
  • Image processing pixel?? (pepper noise)
  • Sensor network neurons die
  • ?????????????????
  • MinMax???? 2????????
  • Player 1 feature removal mechanism
  • Classifier????????????????
  • Player 2 classifier builder
  • ????????(??)?????

4
SVM (hinge loss)????2player game????Player 1
feature removal mechanism
  • worst case???????????(xi)?Loss?????????????
  • ??????? K (feature?????)

Worst case hinge loss
K???????????
Loss max??????
aij ??????(xi)??????????(j)??? (1-ai)??????(xi)?
??????????? ????????????????????
5
SVM (hinge loss)????2player game????Player 2
classifier builder
  • ????????????loss(hwc)????????w????(SVM)

6
Player 1 feature removal mechanism????
  • ????????loss??????????????????

??????????????Relax
si??????????????loss
????
7
?????????FDROP
  • ???? ???
  • ???????
  • ??????O(??????)
  • SVM?O(??????)

8
??? (??????)
  • Multi-class??????
  • Hinge loss?????(log loss??)??????????(challenge)
  • Kernel?????
  • ????????????????(challenge)
  • ?????????????????????
  • ??????a(?????????????)????relax???????????????????

9
????????(1)??????????????
  • ????????
  • loss??????feature?K???????
  • ??????
  • FDROP?different computational class
  • ???? ????????????

K??????????
10
????????(2)????????????
  • FDROP????????????????????????????????
  • ?????????????????????

Clipped from A. Globerson S. Roweis,
Nightmare at Test Time Robust Learning by
Feature Deletion, ICML 2006
??????????FDROP????????????? ?????????????????????
??????????????????
11
???????SVM??????
  • ???????????????SVM????
  • ??????????????????????

Clipped from A. Globerson S. Roweis,
Nightmare at Test Time Robust Learning by
Feature Deletion, ICML 2006
????? ??????????
????? ?????????
??????? ?????????
Write a Comment
User Comments (0)
About PowerShow.com