BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/971
DTSTAMP:20230914T125945Z
SUMMARY:Towards Verifying AI Systems: Testing of Samplers
DESCRIPTION:Speaker: Kuldeep S. Meel (Computer Science Department\nSchool o
 f Computing\nNational University of Singapore\nSingapore)\n\nAbstract: \nA
 bstract: The modern AI systems significantly differ from traditional syste
 ms in their reliance on probabilistic reasoning. We focus on one core comp
 onent of probabilistic reasoning: sampling from a discrete distribution sp
 ecified by a probabilistic model. The widespread usage of heuristics in di
 fferent sampling-based techniques creates a gap between theory and practic
 e: In theory\, heuristics would nullify guarantees while in practice the h
 euristics seem to work well for problems arising from real-world instances
 . Often statistical tests are employed to argue for the quality of distrib
 utions\, but such statistical tests are usually performed on a tiny number
  of samples for which no theoretical guarantees exist for their accuracy. 
 In contrast to testing for deterministic programs\, where one trace is suf
 ficient to prove the existence of a bug\; such is not the case for sampler
 s as one sample is typically not sufficient to prove non-conformity of the
  sampler to the desired distribution.\nThis makes one wonder: whether it i
 s possible to design testing methodology to test whether a sampler under t
 est generates samples close to a given distribution. We will discuss\, to 
 the best of our knowledge\, the first algorithmic framework\, Barbarik\, t
 o test whether the distribution generated is close to the uniform distribu
 tion. In contrast to the sampling techniques that require an exponential o
 r sub-exponential number of samples for sampler whose support can be repre
 sented by n bits\, Barbarik requires samples independent of n. We present 
 a prototype implementation of Barbarik and use it to test three state of t
 he art uniform samplers over the support defined by combinatorial constrai
 nts. Barbarik is able to provide a certificate of uniformity to one sample
 r and demonstrate non-uniformity for the other two samplers (joint work wi
 th Sourav Chakraborty)\nBio: Kuldeep Meel is an Assistant Professor of Com
 puter Science in School of Computing at the National University of Singapo
 re where he holds the Sung Kah Kay Assistant Professorship. He received hi
 s Ph.D. (2017) and M.S. (2014) degree from Rice University\, and B. Tech. 
 (with Honors) degree (2012) in Computer Science and Engineering from India
 n Institute of Technology\, Bombay. His research interests lie at the inte
 rsection of Artificial Intelligence and Formal Methods. He is a recipient 
 of 2019 NRF Fellowship for AI.  His work received the 2018 Ralph Budd Awa
 rd for Best PhD Thesis in Engineering\, 2014 Outstanding Masters Thesis Aw
 ard from Vienna Center of Logic and Algorithms and Best Student Paper Awar
 d at CP 2015.\n
URL:https://www.tcs.tifr.res.in/web/events/971
DTSTART;TZID=Asia/Kolkata:20190620T110000
DTEND;TZID=Asia/Kolkata:20190620T120000
LOCATION:A-201 Seminar Room
END:VEVENT
END:VCALENDAR