BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1554
DTSTAMP:20250603T045458Z
SUMMARY:Robustness in AI Systems: From Statistical Foundations to Language 
 Model Consistency
DESCRIPTION:Speaker: Subho Majumdar (Vijil)\n\nAbstract: \nThe principles o
 f robust estimation\, designed to maintain performance under distributiona
 l deviations and data corruption\, are a cornerstone of reliable statistic
 al inference. Despite their impressive capabilities\, large language model
 s (LLMs) often produce inconsistent outputs when presented with semantical
 ly equivalent inputs\, undermining their reliability in critical applicati
 ons. This inconsistency mirrors classical robustness problems in statistic
 s\, where small perturbations in input can lead to dramatically different 
 outcomes.\n \nI will present two interconnected lines of work in this are
 a. First\, I will talk about robust high-dimensional regression using dens
 ity power divergence\, a generalization of the maximum likelihood techniqu
 e\, and a number of fundamental results on the asymptotics and influence f
 unction of the resulting robust estimators. Second\, I discuss my work on 
 the consistency problem---the equivalent of robustness in LLMs---concerned
  with measuring and minimizing the sensitivity of LLM outputs to input var
 iations through a combination of controlled synthetic data generation and 
 fine-tuning.\n \nThe talk will conclude with a discussion of open problem
 s at the intersection of classical robustness methods and AI consistency\,
  including robust training techniques and influence function analysis\, fo
 r ensuring AI system reliability in high-stakes applications.\nShort Bio:\
 nSubho Majumdar is co-founder and head of AI at Vijil\, a US-based startup
  that helps enterprises build and operate trustworthy AI agents. Previousl
 y\, he was a senior scientist in the security research team at Splunk and 
 the Data Science and AI Research team at AT&T Labs. He has pioneered the u
 se of trustworthy AI methods in multiple companies\, wrote a book\, and fo
 unded multiple nonprofit efforts in this area. He is a recipient of the In
 ternational Indian Statistical Association (IISA) Early Career Award in St
 atistics and Data Sciences. His research interests are on the security and
  reliability of LLMs and statistical machine learning.\n
URL:https://www.tcs.tifr.res.in/web/events/1554
DTSTART;TZID=Asia/Kolkata:20250610T093000
DTEND;TZID=Asia/Kolkata:20250610T103000
LOCATION:via Zoom in A201
END:VEVENT
END:VCALENDAR