BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Influence Functions - Adrian Goldwaser\, Bruno Mlodozeniec\, Runa 
 Eschenhagen\, University of Cambridge
DTSTART:20250305T110000Z
DTEND:20250305T123000Z
UID:TALK229147@talks.cam.ac.uk
CONTACT:120952
DESCRIPTION:When attempting to understand the behaviour of a machine learn
 ing model\, a common question is: how did the training examples contribute
  to a model output? Which examples contributed the most? This can also be 
 framed as a counterfactual question: how would the final model outputs cha
 nge upon removal of some examples from the training set? The goal of train
 ing data attribution (TDA) methods like influence functions\, which will b
 e the subject of this talk\, is to answer precisely this question.\nIn thi
 s talk\, we will give an introduction to influence functions\, discuss cha
 llenges and approaches to scalability\, and give examples of practical app
 lications. We will show that solving the aforementioned data attribution p
 roblem can be extremely useful. It can help identify pernicious data – f
 rom mislabelled examples\, data responsible for undesirable behaviours (e.
 g. profanity or explicit content) through to data poisoning attacks. Influ
 ence functions can help understand memorisation in neural networks\, provi
 ding mitigations to privacy and copyright concerns\, along with fair data 
 valuation. Influence functions can answer the above TDA problem efficientl
 y without retraining\, using only the local information about the training
  loss function around the final model parameters. They have been successfu
 lly used for these tasks for models ranging from 50 billion parameter Larg
 e Language Models to modern diffusion models.
LOCATION:Cambridge University Engineering Department\, CBL Seminar room BE
 4-38.
END:VEVENT
END:VCALENDAR