BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:The challenge of AI errors and the limitations of training accurat
 e and verifiably stable data-driven AI - Ivan Tyukin (King’s College Lon
 don)
DTSTART:20230309T150000Z
DTEND:20230309T160000Z
UID:TALK198034@talks.cam.ac.uk
CONTACT:Matthew Colbrook
DESCRIPTION:Since the seminal work by Szegedy et al. revealing the apparen
 t sensitivity of deep learning classifiers to small adversarial perturbati
 ons of their input data\, the robustness of modern data-driven AI systems 
 has been a widely discussed and broadly debated issue. Among these adversa
 rial perturbations\, there can exist even universal perturbations which tr
 igger the instability of the network for seemingly any input. The presence
  of such instabilities in a tool which is so widely used in applications g
 ives rise to a fundamental question: are these instabilities typical\, and
  to be expected in modern large-scale AI and deep learning models? Moreove
 r\, is it even possible to compute a data-driven AI model which is both ac
 curate and verifiably stable at the same time? \n\nIn the talk\, we will p
 resent and discuss a list of scenarios enabling the formulation of high-le
 vel verifiable criteria for the detection of instabilities in a broad clas
 s of trained models. However\, as we will show too\, major limitations on 
 the pathway to compute accurate and verifiably stable AI from data remain.
  These limitations constitute a fundamental issue around the possibility o
 f building data-driven systems which are indeed accurate and verifiably st
 able. We will discuss potential approaches to alleviate the problem by acc
 epting the inevitability of errors and finding computationally efficient w
 ays to correct them “on-the-job” with given performance guarantees and
  without re-training.\n
LOCATION:Centre for Mathematical Sciences\, MR4
END:VEVENT
END:VCALENDAR