BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Understanding Black-box Predictions via Influence Functions - Pang
  Wei Koh\, Stanford University
DTSTART:20170720T130000Z
DTEND:20170720T140000Z
UID:TALK73531@talks.cam.ac.uk
CONTACT:Adrian Weller
DESCRIPTION:How can we explain the predictions of a black-box model? In th
 is paper\, we use influence functions -- a classic technique from robust s
 tatistics -- to trace a model’s prediction through the learning algorith
 m and back to its training data\, thereby identifying training points most
  responsible for a given prediction. To scale up influence functions to mo
 dern machine learning settings\, we develop a simple\, efficient implement
 ation that requires only oracle access to gradients and Hessian-vector pro
 ducts. We show that even on non-convex and non-differentiable models where
  the theory breaks down\, approximations to influence functions can still 
 provide valuable information. On linear models and convolutional neural ne
 tworks\, we demonstrate that influence functions are useful for multiple p
 urposes: understanding model behavior\, debugging models\, detecting datas
 et errors\, and even creating visually-indistinguishable training-set atta
 cks.
LOCATION:CBL Room BE-438\, Department of Engineering
END:VEVENT
END:VCALENDAR