Transcription of Auxiliary Deep Generative Models - arXiv
{{id}} {{{paragraph}}}
Example: marketing
Auxiliary Deep Generative Models - arXiv
Auxiliary deep Generative ModelsLars Maal Kaae S ren Kaae S of Applied Mathematics and Computer Science, Technical University of Denmark2 Bioinformatics Centre, Department of Biology, University of CopenhagenAbstractDeep Generative Models parameterized by neu-ral networks have recently achieved state-of-the-art performance in unsupervised and semi-supervised learning. We extend deep genera-tive Models with Auxiliary variables which im-proves the variational approximation. The aux-iliary variables leave the Generative model un-changed but make the variational distributionmore expressive. Inspired by the structure of theauxiliary variable we also propose a model withtwo stochastic layers and skip connections. Ourfindings suggest that more expressive and prop-erly specified deep Generative Models convergefaster with better results. We show state-of-the-art performance within semi-supervised learningon MNIST, SVHN and NORB IntroductionStochastic backpropagation, deep neural networks and ap-proximate Bayesian inference have made deep generativemodels practical for large scale problems (Kingma, 2013;Rezende et al.)
Auxiliary Deep Generative Models where a, y, zare the auxiliary variable, class label, and la-tent features, respectively. Cat() is a multinomial distribu-
Loading..
Tags:
Information
Domain:
Source:
Link to this page:
Please notify us if you found a problem with this document: