BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Bias Mitigation in the Wild: Challenges and Opportunities - Mateo 
 Espinosa Zarlenga (University of Cambridge)
DTSTART:20241119T130000Z
DTEND:20241119T140000Z
UID:TALK223087@talks.cam.ac.uk
CONTACT:Mateja Jamnik
DESCRIPTION:Deep neural networks trained via empirical risk minimisation o
 ften exhibit significant performance disparities across groups\, particula
 rly when group and task labels are spuriously correlated (e.g.\, "grassy b
 ackground" and "cows"). In this talk\, I will first argue that previously 
 proposed bias mitigation methods that aim to address this issue often eith
 er rely on group labels for training or validation\, or require an extensi
 ve hyperparameter search\, even when this assumption is not explicitly mad
 e. Such data and computational requirements hinder the practical deploymen
 t of these methods\, especially when datasets are too large to be group-an
 notated\, computational resources are limited\, and models are trained thr
 ough already complex pipelines. With this in mind\, I will outline some of
  the challenges that may need to be addressed to design practical bias mit
 igation methods. Then\, I will describe Targeted Augmentations for Bias mi
 tigation (TAB)\, a new approach that considers these design principles. I 
 will conclude by showing how TAB\, a simple hyperparameter-free framework 
 that leverages the entire training history of a helper model to identify s
 purious samples\, improves worst-group performance without any group infor
 mation or model selection.\n\n"You can also join us on Zoom":https://cam-a
 c-uk.zoom.us/j/83400335522?pwd=LkjYvMOvVpMbabOV1MVTm8QU6DrGN7.1\n
LOCATION:Lecture Theatre 2\, Computer Laboratory\, William Gates Building
END:VEVENT
END:VCALENDAR