Evaluating probabilities under high-dimensional latent variable models

By Iain Murray, Ruslan Salakhutdinov. 2008

91 citations recorded by the source corpus.

Subfields: Probabilistic & Bayesian Methods

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Probabilistic & Bayesian Methods

Citations in the corpus

Deep AutoRegressive Networks (2013, neutral): “DARN with 500 stochastic hidden units compares favourably to state-of-the-art generative performance of deep Boltzmann machines (DBM; Salakhut-dinov & Hinton, 2009) and deep belief networks (DBN; Salakhutdinov & Murray, 2008; Murray & Salakhutdinov, 2009).”
DRAW: A Recurrent Neural Network For Image Generation (2015, methodology): “After a ﬁxed number of glimpses the network uses a softmax layer to classify the MNIST digit.”
Representation Learning: A Review and New Perspectives (2012, background): “sampling [150] to estimate the partition function (and thus”
Importance Weighted Autoencoders (2015, background): “By comparison, deep belief networks achieved log-likelihood of approximately -84.55 nats (Murray & Salakhutdinov, 2009), and deep autoregressive networks achieved log-likelihood of -84.13 nats (Gregor et al., 2014).”
RNADE: The real-valued neural autoregressive density-estimator (2013, background): “limitations of the Gaussian-RBM are well documented [13, 14]: its isotropic conditional noise model does not give competitive density estimates.”
A Deep and Tractable Density Estimator (2013, methodology): “…method with mixtures of multivariate Bernoulli distributions with 10 and 500 components, ﬁxed-ordering NADEs, RBMs (500 hidden units), and two-hidden-layer DBNs (500 and 2000 hidden units on each layer) whose performance was estimated by Salakhutdinov & Murray (2008); Murray & Salakhutdinov (2009).”
The Variational Gaussian Process (2015, background): “Although we haven’t found it necessary in practice, see Appendix E for scaling the size of variational data.”
Discrete Variational Autoencoders (2016, methodology): “In contrast to a DBN, we use a bipartite Boltz-mann machine, with both sides of the bipartite split connected to the rest of the model.”
Reweighted Wake-Sleep (2014, methodology): “We use the MNIST dataset that was binarized according to Murray and Salakhutdinov (2009) and downloaded in it’s binarized form from (Larochelle, 2011).”
Inductive Principles for Restricted Boltzmann Machine Learning (2010, methodology): “We compute test set log likelihoods by estimating the log partition function once for each trained model using the annealed importance sampling (AIS) method proposed by Murray and Salakhutdinov (2009).”
Edinburgh Research Explorer Continuous Relaxations for Discrete Hamiltonian Monte Carlo (unknown): “First, because it is an estimator of Z − 1 , as [15] explains, such an estimator may underestimate Z and log Z in practice.”
Learning generative models of mid-level structure in natural images (2012, background): “Interestingly, several analyses that have been performed suggest that increasing the depth of the hierarchy leads, at best, only to small improvements (e.g. Salakhutdinov and Murray, 2008; Salakhutdinov and Hinton, 2009; Murray and Salakhutdinov, 2009; Theis et al., 2010).”
Bidirectional Helmholtz Machines (2015, methodology): “2 Binarized MNIST Experiments We use the MNIST data set that was binarized according to [70] and downloaded in binarized form [57].”