BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:A Non-Equilibrium Transport Sampler - Eric Vanden-Eijnden (Courant
  Institute of Mathematical Sciences)
DTSTART:20241127T091500Z
DTEND:20241127T100500Z
UID:TALK221533@talks.cam.ac.uk
DESCRIPTION:I will present an algorithm\, termed the Non-Equilibrium Trans
 port Sampler (NETS)\, to sample from unnormalized probability distribution
 s. NETS can be viewed as a variant of annealed importance sampling (AIS) b
 ased on Jarzynski's equality\, in which the stochastic differential equati
 on used to perform the non-equilibrium sampling is augmented with an addit
 ional learned drift term that lowers the impact of the unbiasing weights u
 sed in AIS. This drift is the minimizer of a variety of objective function
 s\, which can all be estimated in an unbiased fashion without backpropagat
 ing through solutions of the stochastic differential equations governing t
 he sampling. I will also prove that some these objectives control the Kull
 back-Leibler divergence of the estimated distribution from its target. NET
 S is shown to be unbiased and\, in addition\, has a tunable diffusion coef
 ficient which can be adjusted post-training to maximize the effective samp
 le size. Tthe efficacy of the method will be illustrated on standard bench
 marks\, high-dimensional Gaussian mixture distributions\, and a model from
  statistical lattice field theory\, for which it surpasses the performance
 s of related work and existing baselines.\nThis is joint work with Michael
  Albergo.
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR