Amortized variational inference (AVI) replaces instance-specific local
inference with a global inference network. While AVI has enabled efficient
training of deep generative models such as variational autoencoders (VAE),
recent empirical work suggests that inference networks can produce suboptimal
variational parameters. We propose a hybrid approach, to use AVI to initialize
the variational parameters and run stochastic variational inference (SVI) to
refine them. Crucially, the local SVI procedure is itself differentiable, so
the inference network and generative model can be trained end-to-end with
gradient-based optimization. This semi-amortized approach enables the use of
rich generative models without experiencing the posterior-collapse phenomenon
common in training VAEs for problems like text generation. Experiments show
this approach outperforms strong autoregressive and variational baselines on
standard text and image datasets.
%0 Generic
%1 kim2018semiamortized
%A Kim, Yoon
%A Wiseman, Sam
%A Miller, Andrew C.
%A Sontag, David
%A Rush, Alexander M.
%D 2018
%K autoencoder unsupervised variational-ae
%T Semi-Amortized Variational Autoencoders
%U http://arxiv.org/abs/1802.02550
%X Amortized variational inference (AVI) replaces instance-specific local
inference with a global inference network. While AVI has enabled efficient
training of deep generative models such as variational autoencoders (VAE),
recent empirical work suggests that inference networks can produce suboptimal
variational parameters. We propose a hybrid approach, to use AVI to initialize
the variational parameters and run stochastic variational inference (SVI) to
refine them. Crucially, the local SVI procedure is itself differentiable, so
the inference network and generative model can be trained end-to-end with
gradient-based optimization. This semi-amortized approach enables the use of
rich generative models without experiencing the posterior-collapse phenomenon
common in training VAEs for problems like text generation. Experiments show
this approach outperforms strong autoregressive and variational baselines on
standard text and image datasets.
@misc{kim2018semiamortized,
abstract = {Amortized variational inference (AVI) replaces instance-specific local
inference with a global inference network. While AVI has enabled efficient
training of deep generative models such as variational autoencoders (VAE),
recent empirical work suggests that inference networks can produce suboptimal
variational parameters. We propose a hybrid approach, to use AVI to initialize
the variational parameters and run stochastic variational inference (SVI) to
refine them. Crucially, the local SVI procedure is itself differentiable, so
the inference network and generative model can be trained end-to-end with
gradient-based optimization. This semi-amortized approach enables the use of
rich generative models without experiencing the posterior-collapse phenomenon
common in training VAEs for problems like text generation. Experiments show
this approach outperforms strong autoregressive and variational baselines on
standard text and image datasets.},
added-at = {2018-02-10T13:06:51.000+0100},
author = {Kim, Yoon and Wiseman, Sam and Miller, Andrew C. and Sontag, David and Rush, Alexander M.},
biburl = {https://www.bibsonomy.org/bibtex/23c14c88ddfe478d6bbe7090dffd156e9/jk_itwm},
description = {Semi-Amortized Variational Autoencoders},
interhash = {eb910cad347567e14d1ee990ebc4b61f},
intrahash = {3c14c88ddfe478d6bbe7090dffd156e9},
keywords = {autoencoder unsupervised variational-ae},
note = {cite arxiv:1802.02550},
timestamp = {2018-02-10T13:06:51.000+0100},
title = {Semi-Amortized Variational Autoencoders},
url = {http://arxiv.org/abs/1802.02550},
year = 2018
}