BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Statistical curriculum learning: An elimination algorithm achievin
 g the weak oracle risk
DTSTART:20240304T161500
DTEND:20240304T170000
DTSTAMP:20260501T120512Z
UID:e8eee3c92bf9d9c24e8be6fdb9cee15a1dd2cb4b67ec88efb3ac66af
CATEGORIES:Conferences - Seminars
DESCRIPTION:Prof. Nir Weinberger\n\nThe Viterbi Faculty of Electrical and 
 Computer Engineering\n\nTechnion - Israel Institute of Technology\nCurricu
 lum Learning (CL) is a widely used machine learning strategy that improves
  the learner's performance by allowing it to order the training samples du
 ring learning\, similarly to the way humans learn. \n\n \n\nIn this work
 \, we address statistical aspects of CL\, and consider a parametric learni
 ng problem with a target task and multiple source tasks. While only the ta
 rget parameter is of interest\, sampling from a source task might be bene
 ficial if they are less noisy than the target task\, while their correspon
 ding parameters are very close. The learner is restricted by the total num
 ber of samples\, and can adaptively choose how to allocate samples to each
  of the models. \n\n \n\nWe define a strong-oracle learner as an ideal l
 earner\, which allocates all its samples to the most effective model (eith
 er the target or one of the sources). We show that achieving its performan
 ce is too ambitious for a learning algorithm\, and advocate a weak-oracle
  learner\, as a more realistic benchmark for CL algorithms. \n\n \n\nWe 
 first develop an elimination-based learning algorithm\, and determine con
 ditions that allow it to match the weak-oracle learner. We then consider
  lower bounds via minimax lower bounds.  We reveal a few challenges assoc
 iated with defining informative classes of problem-instances\, propose two
  bounds\, and determine the conditions under which the performance weak-or
 acle learner is provably optimal.\n\n \n\nJoint work with Omer Cohen and 
 Ron Meir. 
LOCATION:BC 129 https://plan.epfl.ch/?room==BC%20129
STATUS:CANCELLED
END:VEVENT
END:VCALENDAR