BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Pitfalls with ablation in neural network architectures - Christina
  Lioma (University of Copenhagen)
DTSTART:20220617T110000Z
DTEND:20220617T120000Z
UID:TALK175817@talks.cam.ac.uk
CONTACT:Michael Schlichtkrull
DESCRIPTION:Abstract:\n\nAblation tests are frequently used for analysing 
 machine learning performance. Generally\, ablation refers to the removal o
 f a component in order to understand its contribution to the overall decis
 ion-making process. For instance\, it is expected that the ablation of a f
 eature during classification will affect the output somehow analogously to
  the importance of that feature for the classification task at hand. Ablat
 ion is therefore routinely used to attribute feature importance\, as well 
 as to explain machine learning output in a partial\, approximate\, and mod
 el-agnostic way.\n\nThis talk will point out a core problem when using abl
 ation with neural network architectures. The problem stems from the tenden
 cy of neural network architectures to ignore complex predictive features i
 n the presence of few simple predictive features\, even when the complex f
 eatures have significantly greater predictive power than the simple featur
 es. This talk will provide evidence demonstrating the existence of this te
 ndency\, even in small neural network architectures\, and show how this ma
 y entirely invalidate the standard interpretation of ablation tests. A dis
 cussion about why this is important and ways of moving forward will be pro
 vided.\n\nThis is joint work with Qiuchi Li (University of Copenhagen).\n\
 nBio:\n\nChristina Lioma is a professor in machine learning at the Departm
 ent of Computer Science\, University of Copenhagen. Her research focuses o
 n applied machine learning\, information retrieval and web search technolo
 gies\, web data mining and analytics\, recommendation systems and natural 
 language processing. She has a track-record of research collaboration with
  Danish and international industry\, and an alumni of >20 PhD students and
  postdocs. Since 2012\, Christina Lioma has attracted more than 50 million
  USD in external funding.
LOCATION:Computer Lab\, SS03
END:VEVENT
END:VCALENDAR