BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
X-WR-CALNAME:Challenges in AI-assisted Supervision of AIs
X-WR-TIMEZONE:Eastern Time (US & Canada)
BEGIN:VEVENT
DTSTAMP:20261013T013646Z
UID:tag:localist.com\,2008:EventInstance_47385471201865
DTSTART:20240923T193000Z
DTEND:20240923T210000Z
DESCRIPTION:Please join the GW Co-Design of Trustworthy AI in Systems (DTAI
 S) program for this "advanced intro" seminar!\n\nThe growing capability of
  AIs motivated us to use them in more and more complex tasks. At the same 
 time\, AIs frequently make subtle mistakes that are difficult for humans t
 o identify. It becomes challenging to get reliable human supervision of AI
 s\, including capability assessment\, evaluation of individual predictions
 \, and effective delegation. It seems inevitable that we will need to enli
 st the help of AIs to help us supervise AIs. In this talk\, Professor Feng
  will start with a high-level overview of the technical alignment and AI s
 afety research landscape\, then discuss challenges in using AIs to supervi
 se themselves. The discussion will cover three specific projects on how se
 lf-evaluation can be biased and methods for assisting human evaluation can
  have the opposite effect. The first projects studies language models' sel
 f-recognition capability and how it can bias self-evaluation\; the second 
 project studies iterative self-refinement and how it can cause LMs to furt
 her diverge from human preferences\; the third project studies how RLHF fr
 om flawed human feedback can make models more deceptive.\n\nAttendance is 
 virtual and the content is suitable for TAI experts and newcomers alike.
LOCATION:
SUMMARY:Challenges in AI-assisted Supervision of AIs
URL;VALUE=URI:https://calendar.gwu.edu/event/LLMs-and-explainability-prof-f
 eng
CATEGORIES:Lectures & Speakers
CATEGORIES:Science & Technology
END:VEVENT
END:VCALENDAR