BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Interpretability - the myth\, questions\, and some answers - Been 
 Kim\, Google Brain
DTSTART:20180910T100000Z
DTEND:20180910T110000Z
UID:TALK109048@talks.cam.ac.uk
CONTACT:Adrian Weller
DESCRIPTION:NOTE LOCATION: 2ND FLOOR BOARD ROOM\n\nIn this talk\, I will p
 rovide an overview of my work on interpretability from the past couple of 
 years. I will talk about 1) our studies on factors that influence how huma
 ns understand explanations from machine learning models\, 2) building inhe
 rently interpretable models with and without human-in-the-loop\, 3) improv
 ing interpretability when you already have a model (post-training interpre
 tability) and 4) our work on ways to test and evaluate interpretability me
 thods. \n\nAmong them\, I will take a deeper dive in one of my recent work
 s - testing with concept activation vectors (TCAV) - a post-training inter
 pretability method for complex models\, such as neural networks. This meth
 od provides an interpretation of a neural net's internal state in terms of
  human-friendly\, high-level concepts instead of low-level input features.
  The key idea is to view the high-dimensional internal state of a neural n
 et as an aid\, not an obstacle. We show how to use concept activation vect
 ors (CAVs) as part of a technique\, Testing with CAVs (TCAV)\, that uses d
 irectional derivatives to quantify the degree to which a user-defined conc
 ept is important to a classification result--for example\, how sensitive a
  prediction of “zebra” is to the presence of stripes. Using the domain
  of image classification as a testing ground\, we describe how CAVs may be
  used to explore hypotheses and generate insights for a standard image cla
 ssification network as well as a medical application.\n
LOCATION:Engineering Department\, Board Room on the 2nd Floor
END:VEVENT
END:VCALENDAR