BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Cambridge Ellis Unit
SUMMARY:Cambridge ELLIS Seminar Series- Dr Roger Grosse- S
 tudying Neural Net Generalization through Influenc
 e Functions - Speaker to be confirmed
DTSTART;TZID=Europe/London:20230306T161500
DTEND;TZID=Europe/London:20230306T171500
UID:TALK197914AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/197914
DESCRIPTION:The Cambridge ELLIS Unit Seminar Series holds talk
 s by leading researchers in the area of machine le
 arning and AI. Next first  speaker of 2023 will be
  Dr. Roger Grosse. Details of his talk can be foun
 d below. \n\nTitle: “Studying Neural Net Generaliz
 ation through Influence Functions"\n\nAbstract: Ho
 w can we trace surprising behaviors of machine lea
 rning models back to their training data?  Influen
 ce functions aim to predict how the trained model 
 would change if a specific training example were a
 dded to the training set. I’ll address two issues 
 that have blocked their applicability to large-sca
 le neural nets: apparent inaccuracy of the results
 \, and the difficulty of computing inverse-Hessian
 -vector products. Towards the former issue\, I’ll 
 reformulate the goals of influence estimation in a
  way that applies to overparameterized\, incomplet
 ely trained models\, and argue that the apparent i
 naccuracy was largely illusory. I’ll then discuss 
 an approach to scaling influence estimation to lar
 ge language models and show some resulting insight
 s into their patterns of generalization..\n\nhttp
 s://eng-cam.zoom.us/j/82557742057?pwd=Sk1sRWRxL09p
 V2FPYzNpK0FrU096Zz09\n
LOCATION:https://eng-cam.zoom.us/j/82557742057?pwd=Sk1sRWRx
 L09pV2FPYzNpK0FrU096Zz09
CONTACT:Kimberly Cole
END:VEVENT
END:VCALENDAR