BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Classification with imperfect training labels - Timothy Cannings (
 University of Southern California)
DTSTART:20180607T100000Z
DTEND:20180607T110000Z
UID:TALK107260@talks.cam.ac.uk
CONTACT:INI IT
DESCRIPTION:We study the effect of imperfect training data labels on the p
 erformance of  classification methods.  In a general setting\, where the p
 robability that an observation  in the training dataset is mislabelled may
  depend on both the feature vector and the true label\, we bound the exces
 s risk of an arbitrary classifier trained with imperfect labels in terms o
 f its excess risk for predicting a noisy label. This reveals conditions un
 der which a classifier trained with imperfect labels remains consistent fo
 r classifying uncorrupted test data points. Furthermore\, under stronger c
 onditions\, we derive detailed asymptotic properties for the popular k-nea
 rest neighbour (k-nn)\, Support Vector Machine (SVM) and Linear Discrimina
 nt Analysis (LDA) classifiers. One consequence of these results is that th
 e k-nn and SVM classifiers are robust to imperfect training labels\, in th
 e sense that the rate of convergence of the excess risks of these classifi
 ers remains unchanged\; in fact\, it even turns out that in some cases\, i
 mperfect labels may improve the performance of these methods.  On the othe
 r hand\, the LDA classifier is shown to be typically inconsistent in the p
 resence of label noise unless the prior probabilities of each class are eq
 ual.
LOCATION:Seminar Room 2\, Newton Institute
END:VEVENT
END:VCALENDAR