The kernel exponential family is a rich class of distributions,which can be
fit efficiently and with statistical guarantees by score matching. Being
required to choose a priori a simple kernel such as the Gaussian, however,
limits its practical applicability. We provide a scheme for learning a kernel
parameterized by a deep network, which can find complex location-dependent
local features of the data geometry. This gives a very rich class of density
models, capable of fitting complex structures on moderate-dimensional problems.
Compared to deep density models fit via maximum likelihood, our approach
provides a complementary set of strengths and tradeoffs: in empirical studies,
the former can yield higher likelihoods, whereas the latter gives better
estimates of the gradient of the log density, the score, which describes the
distribution's shape.
Beschreibung
[1811.08357] Learning deep kernels for exponential family densities
%0 Journal Article
%1 wenliang2018learning
%A Wenliang, Li
%A Sutherland, Dougal
%A Strathmann, Heiko
%A Gretton, Arthur
%D 2018
%K bayesian kernels probability readings stats
%T Learning deep kernels for exponential family densities
%U http://arxiv.org/abs/1811.08357
%X The kernel exponential family is a rich class of distributions,which can be
fit efficiently and with statistical guarantees by score matching. Being
required to choose a priori a simple kernel such as the Gaussian, however,
limits its practical applicability. We provide a scheme for learning a kernel
parameterized by a deep network, which can find complex location-dependent
local features of the data geometry. This gives a very rich class of density
models, capable of fitting complex structures on moderate-dimensional problems.
Compared to deep density models fit via maximum likelihood, our approach
provides a complementary set of strengths and tradeoffs: in empirical studies,
the former can yield higher likelihoods, whereas the latter gives better
estimates of the gradient of the log density, the score, which describes the
distribution's shape.
@article{wenliang2018learning,
abstract = {The kernel exponential family is a rich class of distributions,which can be
fit efficiently and with statistical guarantees by score matching. Being
required to choose a priori a simple kernel such as the Gaussian, however,
limits its practical applicability. We provide a scheme for learning a kernel
parameterized by a deep network, which can find complex location-dependent
local features of the data geometry. This gives a very rich class of density
models, capable of fitting complex structures on moderate-dimensional problems.
Compared to deep density models fit via maximum likelihood, our approach
provides a complementary set of strengths and tradeoffs: in empirical studies,
the former can yield higher likelihoods, whereas the latter gives better
estimates of the gradient of the log density, the score, which describes the
distribution's shape.},
added-at = {2020-02-12T21:41:26.000+0100},
author = {Wenliang, Li and Sutherland, Dougal and Strathmann, Heiko and Gretton, Arthur},
biburl = {https://www.bibsonomy.org/bibtex/27c538f0c5004df4cbf401954503528dc/kirk86},
description = {[1811.08357] Learning deep kernels for exponential family densities},
interhash = {ada37bddec0d1d7a74b364f548ab3054},
intrahash = {7c538f0c5004df4cbf401954503528dc},
keywords = {bayesian kernels probability readings stats},
note = {cite arxiv:1811.08357},
timestamp = {2020-02-12T21:41:26.000+0100},
title = {Learning deep kernels for exponential family densities},
url = {http://arxiv.org/abs/1811.08357},
year = 2018
}