BEGIN:VCALENDAR
VERSION:2.0
PRODID:www.imperial.ac.uk
BEGIN:VEVENT
UID:63d72deb68288
DTSTART:20210212T140000Z
SEQUENCE:0
TRANSP:OPAQUE
DTEND:20210212T150000Z
URL:https://www.imperial.ac.uk/events/129770/statistics-seminar-dr-ritabrat
a-dutta-warwick-score-matched-conditional-exponential-families-for-likelih
ood-free-inference/
LOCATION:United Kingdom
SUMMARY:Statistics Seminar: Dr Ritabrata Dutta (Warwick) — Score Matched
Conditional Exponential Families for Likelihood-Free Inference
CLASS:PUBLIC
DESCRIPTION:Title: Score Matched Conditional Exponential Families for Likel
ihood-Free Inference\nAbstract: To perform Bayesian inference for stochast
ic simulator models for which the likelihood is not accessible\, Likelihoo
d-Free Inference (LFI) relies on simulations from the model. Standard LFI
methods can be split according to how these simulations are used: to build
an explicit Surrogate Likelihood\, or to accept/reject parameter values a
ccording to a measure of distance from the observations (Approximate Bayes
ian Computation (ABC)). In both cases\, simulations are adaptively tailore
d to the value of the observation. Here\, we generate parameter-simulation
pairs from the model independently on the observation\, and use them to l
earn a conditional exponential family likelihood approximation\; to parame
trize it\, we use Neural Networks whose weights are tuned with Score Match
ing. With our likelihood approximation\, we can employ MCMC for doubly int
ractable distributions to draw samples from the posterior for any number o
f observations without additional model simulations\, with performance com
petitive to comparable approaches. Further\, the sufficient statistics of
the exponential family can be used as summaries in ABC\, outperforming the
state-of-the-art method in five different models with known likelihood. F
inally\, we apply our method to a challenging model from meteorology.
X-ALT-DESC;FMTTYPE=text/html:**Title**: Score Matched Conditional Exp
onential Families for Likelihood-Free Inference

\n**Abstract**: T
o perform Bayesian inference for stochastic simulator models for which the
likelihood is not accessible\, Likelihood-Free Inference (LFI) relies on
simulations from the model. Standard LFI methods can be split according to
how these simulations are used: to build an explicit Surrogate Likelihood
\, or to accept/reject parameter values according to a measure of distance
from the observations (Approximate Bayesian Computation (ABC)). In both c
ases\, simulations are adaptively tailored to the value of the observation
. Here\, we generate parameter-simulation pairs from the model independent
ly on the observation\, and use them to learn a conditional exponential fa
mily likelihood approximation\; to parametrize it\, we use Neural Networks
whose weights are tuned with Score Matching. With our likelihood approxim
ation\, we can employ MCMC for doubly intractable distributions to draw sa
mples from the posterior for any number of observations without additional
model simulations\, with performance competitive to comparable approaches
. Further\, the sufficient statistics of the exponential family can be use
d as summaries in ABC\, outperforming the state-of-the-art method in five
different models with known likelihood. Finally\, we apply our method to a
challenging model from meteorology.

DTSTAMP:20230130T023939Z
END:VEVENT
END:VCALENDAR