Y. Yang, R. Bamler, and S. Mandt. (2020)cite arxiv:2006.04240Comment: 8 pages + detailed supplement with additional qualitative and quantitative results.
Abstract
We consider the problem of lossy image compression with deep latent variable
models. State-of-the-art methods build on hierarchical variational
autoencoders~(VAEs) and learn inference networks to predict a compressible
latent representation of each data point. Drawing on the variational inference
perspective on compression, we identify three approximation gaps which limit
performance in the conventional approach: (i)~an amortization gap, (ii)~a
discretization gap, and (iii)~a marginalization gap. We propose improvements to
each of these three shortcomings based on iterative inference, stochastic
annealing for discrete optimization, and bits-back coding, resulting in the
first application of bits-back coding to lossy compression. In our experiments,
which include extensive baseline comparisons and ablation studies, we achieve
new state-of-the-art performance on lossy image compression using an
established VAE architecture, by changing only the inference method.
Description
[2006.04240] Improving Inference for Neural Image Compression
%0 Journal Article
%1 yang2020improving
%A Yang, Yibo
%A Bamler, Robert
%A Mandt, Stephan
%D 2020
%K compression inference variational
%T Improving Inference for Neural Image Compression
%U http://arxiv.org/abs/2006.04240
%X We consider the problem of lossy image compression with deep latent variable
models. State-of-the-art methods build on hierarchical variational
autoencoders~(VAEs) and learn inference networks to predict a compressible
latent representation of each data point. Drawing on the variational inference
perspective on compression, we identify three approximation gaps which limit
performance in the conventional approach: (i)~an amortization gap, (ii)~a
discretization gap, and (iii)~a marginalization gap. We propose improvements to
each of these three shortcomings based on iterative inference, stochastic
annealing for discrete optimization, and bits-back coding, resulting in the
first application of bits-back coding to lossy compression. In our experiments,
which include extensive baseline comparisons and ablation studies, we achieve
new state-of-the-art performance on lossy image compression using an
established VAE architecture, by changing only the inference method.
@article{yang2020improving,
abstract = {We consider the problem of lossy image compression with deep latent variable
models. State-of-the-art methods build on hierarchical variational
autoencoders~(VAEs) and learn inference networks to predict a compressible
latent representation of each data point. Drawing on the variational inference
perspective on compression, we identify three approximation gaps which limit
performance in the conventional approach: (i)~an amortization gap, (ii)~a
discretization gap, and (iii)~a marginalization gap. We propose improvements to
each of these three shortcomings based on iterative inference, stochastic
annealing for discrete optimization, and bits-back coding, resulting in the
first application of bits-back coding to lossy compression. In our experiments,
which include extensive baseline comparisons and ablation studies, we achieve
new state-of-the-art performance on lossy image compression using an
established VAE architecture, by changing only the inference method.},
added-at = {2020-06-09T09:28:03.000+0200},
author = {Yang, Yibo and Bamler, Robert and Mandt, Stephan},
biburl = {https://www.bibsonomy.org/bibtex/26adb4ef2be36b803f38b997a81e79720/kirk86},
description = {[2006.04240] Improving Inference for Neural Image Compression},
interhash = {d3174d993212808be2ddf6eb6a0c668b},
intrahash = {6adb4ef2be36b803f38b997a81e79720},
keywords = {compression inference variational},
note = {cite arxiv:2006.04240Comment: 8 pages + detailed supplement with additional qualitative and quantitative results},
timestamp = {2020-06-09T09:28:03.000+0200},
title = {Improving Inference for Neural Image Compression},
url = {http://arxiv.org/abs/2006.04240},
year = 2020
}