BEGIN:VCALENDAR
PRODID:-//eluceo/ical//2.0/EN
VERSION:2.0
CALSCALE:GREGORIAN
BEGIN:VEVENT
UID:www.tcs.tifr.res.in/event/1717
DTSTAMP:20260508T090820Z
SUMMARY:Adversarial Hypothesis Testing: Channel Estimation\, Sequentiality 
 and Robustness
DESCRIPTION:Speaker: Eeshan Modak (TIFR)\n\nAbstract: \nIn this thesis\, we
  study the following problems:\nAdversarial hypothesis testing is a model 
 for problems where the observed data is not independent and identically di
 stributed according to a fixed distribution. The samples could instead com
 e from distributions arbitrarily chosen by an adversary. We show how seque
 ntial tests can obtain a strictly better performance compared to fixed len
 gth tests in this setting.\nArbitrarily Varying Channels (AVC's) model cha
 nnels which can vary with time in an arbitrary way during the transmission
 . We study the problem of distinguishing between two AVC's where the trans
 mitter (i) is deterministic\, (ii) may privately randomize\, and (iii) sha
 res randomness with the detector.\nIn many practical hypothesis testing pr
 oblems\, our hypotheses might not exactly model the observed data. In such
  a situation\, we would like our test to output the hypothesis which is cl
 oser to the true distribution of the underlying data. It turns out that th
 is is possible only when the hypotheses are not too close. We give a lower
  bound on the optimal separation when the closeness is measured in terms o
 f the Hellinger distance. \nObtaining bounds on the expected generalizati
 on error of a machine learning algorithm is an important problem. We obtai
 n a family of Rényi divergence-based bounds that recover some of the exis
 ting bounds as a special case. Also\, for certain values of the Rényi par
 ameter\, they can be tighter than the existing bounds.\n
URL:https://www.tcs.tifr.res.in/web/events/1717
DTSTART;TZID=Asia/Kolkata:20260529T103000
DTEND;TZID=Asia/Kolkata:20260529T113000
LOCATION:A-201 (STCS Seminar Room)
END:VEVENT
END:VCALENDAR