BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning Backward Compatible Embeddings - Weihua Hu\, Stanford Uni
 versity and Google
DTSTART:20220902T150000Z
DTEND:20220902T160000Z
UID:TALK178034@talks.cam.ac.uk
CONTACT:Pietro Lio
DESCRIPTION:Embeddings\, low-dimensional vector representation of objects\
 , are fundamental in building modern machine learning systems. In industri
 al settings\, there is usually an embedding team that trains an embedding 
 model to solve intended tasks (e.g.\, product recommendation). The produce
 d embeddings are then widely consumed by consumer teams to solve their uni
 ntended tasks (e.g.\, fraud detection). However\, as the embedding model g
 ets updated and retrained to improve performance on the intended task\, th
 e newly-generated embeddings are no longer compatible with the existing co
 nsumer models. Here we study the problem of embedding version updates and 
 their backward compatibility. We formalize the problem where the goal is f
 or the embedding team to keep updating the embedding version\, while the c
 onsumer teams do not have to retrain their models. We develop a solution b
 ased on learning backward compatible embeddings\, which allows the embeddi
 ng model version to be updated frequently\, while also allowing the latest
  version of the embedding to be quickly transformed into any backward comp
 atible historical version of it\, so that consumer teams do not have to re
 train their models. We explore different design choices under our framewor
 k and show effectiveness of our approach.
LOCATION:Lecture Theatre 2 (https://cl-cam-ac-uk.zoom.us/j/92564083880?pwd
 =Wk5DVWpUc2lIN0krMFU5azEwUGpEUT09)
END:VEVENT
END:VCALENDAR