BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Inverting Language Models - Alexander (Sasha) Rush\, Cornell and H
 ugging Face
DTSTART:20240314T140000Z
DTEND:20240314T150000Z
UID:TALK213124@talks.cam.ac.uk
CONTACT:Panagiotis Fytas
DESCRIPTION:As language models enter production environments\, their inter
 mediate states are used for a myriad of downstream applications such as se
 arch\, prompting\, and document comparison. In this talk\, I discuss the f
 easibility of language model inversion. Specifically\, we are interested i
 n how much information language models contain about their inputs? We inve
 stigate the problem in two scenarios\, recovering text inputs from the out
 puts of embeddings from sentence embedders and next-token probability outp
 uts from language models. In many cases\, our methods are able to fully re
 cover exact textual inputs given just intermediate states. I'll discuss th
 e security implications of these findings\, as well as what this tells us 
 about compression embedding and language modeling applications. 
LOCATION:https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBd
 XVpOXFvdz09
END:VEVENT
END:VCALENDAR