variational autoencoder

{{short description|Deep learning generative model to encode data representation}}

In machine learning, a variational autoencoder (VAE) is an artificial neural network architecture introduced by Diederik P. Kingma and Max Welling.{{cite arXiv |last1=Kingma |first1=Diederik P. |title=Auto-Encoding Variational Bayes |date=2022-12-10 |last2=Welling |first2=Max|class=stat.ML |eprint=1312.6114 }} It is part of the families of probabilistic graphical models and variational Bayesian methods.{{cite book |first1=Lucas |last1=Pinheiro Cinelli |first2=Matheus |last2=Araújo Marins |first3=Eduardo Antônio |last3=Barros da Silva |first4=Sérgio |last4=Lima Netto |display-authors=1 |title=Variational Methods for Machine Learning with Applications to Deep Networks |location= |publisher=Springer |year=2021 |pages=111–149 |chapter=Variational Autoencoder |isbn=978-3-030-70681-4 |chapter-url=https://books.google.com/books?id=N5EtEAAAQBAJ&pg=PA111 |doi=10.1007/978-3-030-70679-1_5 |s2cid=240802776 }}

In addition to being seen as an autoencoder neural network architecture, variational autoencoders can also be studied within the mathematical formulation of variational Bayesian methods, connecting a neural encoder network to its decoder through a probabilistic latent space (for example, as a multivariate Gaussian distribution) that corresponds to the parameters of a variational distribution.

Thus, the encoder maps each point (such as an image) from a large complex dataset into a distribution within the latent space, rather than to a single point in that space. The decoder has the opposite function, which is to map from the latent space to the input space, again according to a distribution (although in practice, noise is rarely added during the decoding stage). By mapping a point to a distribution instead of a single point, the network can avoid overfitting the training data. Both networks are typically trained together with the usage of the reparameterization trick, although the variance of the noise model can be learned separately.{{cn|date=June 2024}}

Although this type of model was initially designed for unsupervised learning,{{cite arXiv |last1=Dilokthanakul |first1=Nat |last2=Mediano |first2=Pedro A. M. |last3=Garnelo |first3=Marta |last4=Lee |first4=Matthew C. H. |last5=Salimbeni |first5=Hugh |last6=Arulkumaran |first6=Kai |last7=Shanahan |first7=Murray |title=Deep Unsupervised Clustering with Gaussian Mixture Variational Autoencoders |date=2017-01-13 |class=cs.LG |eprint=1611.02648}}{{cite book |last1=Hsu |first1=Wei-Ning |last2=Zhang |first2=Yu |last3=Glass |first3=James |title=2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) |chapter=Unsupervised domain adaptation for robust speech recognition via variational autoencoder-based data augmentation |date=December 2017 |pages=16–23 |doi=10.1109/ASRU.2017.8268911 |arxiv=1707.06265 |isbn=978-1-5090-4788-8 |s2cid=22681625 |chapter-url=https://ieeexplore.ieee.org/document/8268911}} its effectiveness has been proven for semi-supervised learning{{cite book |last1=Ehsan Abbasnejad |first1=M. |last2=Dick |first2=Anthony |last3=van den Hengel |first3=Anton |title=Infinite Variational Autoencoder for Semi-Supervised Learning |date=2017 |pages=5888–5897 |url=https://openaccess.thecvf.com/content_cvpr_2017/html/Abbasnejad_Infinite_Variational_Autoencoder_CVPR_2017_paper.html}}{{cite journal |last1=Xu |first1=Weidi |last2=Sun |first2=Haoze |last3=Deng |first3=Chao |last4=Tan |first4=Ying |title=Variational Autoencoder for Semi-Supervised Text Classification |journal=Proceedings of the AAAI Conference on Artificial Intelligence |date=2017-02-12 |volume=31 |issue=1 |doi=10.1609/aaai.v31i1.10966 |s2cid=2060721 |url=https://ojs.aaai.org/index.php/AAAI/article/view/10966 |language=en|doi-access=free }} and supervised learning.{{cite journal |last1=Kameoka |first1=Hirokazu |last2=Li |first2=Li |last3=Inoue |first3=Shota |last4=Makino |first4=Shoji |title=Supervised Determined Source Separation with Multichannel Variational Autoencoder |journal=Neural Computation |date=2019-09-01 |volume=31 |issue=9 |pages=1891–1914 |doi=10.1162/neco_a_01217 |pmid=31335290 |s2cid=198168155 |url=https://direct.mit.edu/neco/article/31/9/1891/8494/Supervised-Determined-Source-Separation-with|url-access=subscription }}

Overview of architecture and operation

A variational autoencoder is a generative model with a prior and noise distribution respectively. Usually such models are trained using the expectation-maximization meta-algorithm (e.g. probabilistic PCA, (spike & slab) sparse coding). Such a scheme optimizes a lower bound of the data likelihood, which is usually computationally intractable, and in doing so requires the discovery of q-distributions, or variational posteriors. These q-distributions are normally parameterized for each individual data point in a separate optimization process. However, variational autoencoders use a neural network as an amortized approach to jointly optimize across data points. In that way, the same parameters are reused for multiple data points, which can result in massive memory savings. The first neural network takes as input the data points themselves, and outputs parameters for the variational distribution. As it maps from a known input space to the low-dimensional latent space, it is called the encoder.

The decoder is the second neural network of this model. It is a function that maps from the latent space to the input space, e.g. as the means of the noise distribution. It is possible to use another neural network that maps to the variance, however this can be omitted for simplicity. In such a case, the variance can be optimized with gradient descent.

To optimize this model, one needs to know two terms: the "reconstruction error", and the Kullback–Leibler divergence (KL-D). Both terms are derived from the free energy expression of the probabilistic model, and therefore differ depending on the noise distribution and the assumed prior of the data, here referred to as p-distribution. For example, a standard VAE task such as IMAGENET is typically assumed to have a gaussianly distributed noise; however, tasks such as binarized MNIST require a Bernoulli noise. The KL-D from the free energy expression maximizes the probability mass of the q-distribution that overlaps with the p-distribution, which unfortunately can result in mode-seeking behaviour. The "reconstruction" term is the remainder of the free energy expression, and requires a sampling approximation to compute its expectation value.{{cite arXiv |last1=Kingma |first1=Diederik P. |last2=Welling |first2=Max |title=Auto-Encoding Variational Bayes |date=2013-12-20 |class=stat.ML |eprint=1312.6114}}

More recent approaches replace Kullback–Leibler divergence (KL-D) with various statistical distances, see "Statistical distance VAE variants" below.

Formulation

From the point of view of probabilistic modeling, one wants to maximize the likelihood of the data $x$ by their chosen parameterized probability distribution $p_{\theta}(x) = p(x|\theta)$ . This distribution is usually chosen to be a Gaussian $N(x|\mu,\sigma)$ which is parameterized by $\mu$ and $\sigma$ respectively, and as a member of the exponential family it is easy to work with as a noise distribution. Simple distributions are easy enough to maximize, however distributions where a prior is assumed over the latents $z$ results in intractable integrals. Let us find $p_\theta(x)$ via marginalizing over $z$ .

: $p_\theta(x) = \int_{z}p_\theta({x,z}) \, dz,$

where $p_\theta({x,z})$ represents the joint distribution under $p_\theta$ of the observable data $x$ and its latent representation or encoding $z$ . According to the chain rule, the equation can be rewritten as

: $p_\theta(x) = \int_{z}p_\theta({x| z})p_\theta(z) \, dz$

In the vanilla variational autoencoder, $z$ is usually taken to be a finite-dimensional vector of real numbers, and $p_\theta({x|z})$ to be a Gaussian distribution. Then $p_\theta(x)$ is a mixture of Gaussian distributions.

It is now possible to define the set of the relationships between the input data and its latent representation as

Prior $p_\theta(z)$
Likelihood $p_\theta(x|z)$
Posterior $p_\theta(z|x)$

Unfortunately, the computation of $p_\theta(z|x)$ is expensive and in most cases intractable. To speed up the calculus to make it feasible, it is necessary to introduce a further function to approximate the posterior distribution as

: $q_\phi({z| x}) \approx p_\theta({z| x})$

with $\phi$ defined as the set of real values that parametrize $q$ . This is sometimes called amortized inference, since by "investing" in finding a good $q_\phi$ , one can later infer $z$ from $x$ quickly without doing any integrals.

In this way, the problem is to find a good probabilistic autoencoder, in which the conditional likelihood distribution $p_\theta(x|z)$ is computed by the probabilistic decoder, and the approximated posterior distribution $q_\phi(z|x)$ is computed by the probabilistic encoder.

Parametrize the encoder as $E_\phi$ , and the decoder as $D_\theta$ .

Evidence lower bound (ELBO)

Like many deep learning approaches that use gradient-based optimization, VAEs require a differentiable loss function to update the network weights through backpropagation.

For variational autoencoders, the idea is to jointly optimize the generative model parameters $\theta$ to reduce the reconstruction error between the input and the output, and $\phi$ to make $q_\phi({z| x})$ as close as possible to $p_\theta(z|x)$ . As reconstruction loss, mean squared error and cross entropy are often used.

As distance loss between the two distributions the Kullback–Leibler divergence $D_{KL}(q_\phi({z| x})\parallel p_\theta({z| x}))$ is a good choice to squeeze $q_\phi({z| x})$ under $p_\theta(z|x)$ .{{cite news |title=From Autoencoder to Beta-VAE |url=https://lilianweng.github.io/lil-log/2018/08/12/from-autoencoder-to-beta-vae.html |website=Lil'Log |language=en |date=2018-08-12}}

The distance loss just defined is expanded as

: $\begin{align}
D_{KL}(q_\phi({z| x})\parallel p_\theta({z| x})) &= \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{q_\phi(z|x)}{p_\theta(z|x)}\right]\\
&= \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{q_\phi({z| x})p_\theta(x)}{p_\theta(x, z)}\right]\\
&=\ln p_\theta(x) + \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{q_\phi({z| x})}{p_\theta(x, z)}\right]
\end{align}$

Now define the evidence lower bound (ELBO): $L_{\theta,\phi}(x) :=
\mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{p_\theta(x, z)}{q_\phi({z| x})}\right]
= \ln p_\theta(x) - D_{KL}(q_\phi({\cdot| x})\parallel p_\theta({\cdot | x}))$ Maximizing the ELBO $\theta^*,\phi^* = \underset{\theta,\phi}\operatorname{arg max} \, L_{\theta,\phi}(x)$ is equivalent to simultaneously maximizing $\ln p_\theta(x)$ and minimizing $D_{KL}(q_\phi({z| x})\parallel p_\theta({z| x}))$ . That is, maximizing the log-likelihood of the observed data, and minimizing the divergence of the approximate posterior $q_\phi(\cdot | x)$ from the exact posterior $p_\theta(\cdot | x)$ .

The form given is not very convenient for maximization, but the following, equivalent form, is: $L_{\theta,\phi}(x) = \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln p_\theta(x|z)\right] - D_{KL}(q_\phi({\cdot| x})\parallel p_\theta(\cdot))$ where $\ln p_\theta(x|z)$ is implemented as $-\frac{1}{2}\| x - D_\theta(z)\|^2_2$ , since that is, up to an additive constant, what $x|z \sim \mathcal N(D_\theta(z), I)$ yields. That is, we model the distribution of $x$ conditional on $z$ to be a Gaussian distribution centered on $D_\theta(z)$ . The distribution of $q_\phi(z |x)$ and $p_\theta(z)$ are often also chosen to be Gaussians as $z|x \sim \mathcal N(E_\phi(x), \sigma_\phi(x)^2I)$ and $z \sim \mathcal N(0, I)$ , with which we obtain by the formula for KL divergence of Gaussians: $L_{\theta,\phi}(x) = -\frac 12\mathbb E_{z \sim q_\phi(\cdot | x)} \left[ \|x - D_\theta(z)\|_2^2\right] - \frac 12 \left( N\sigma_\phi(x)^2 + \|E_\phi(x)\|_2^2 - 2N\ln\sigma_\phi(x) \right) + Const$ Here $N$ is the dimension of $z$ . For a more detailed derivation and more interpretations of ELBO and its maximization, see its main page.

Reparameterization

File:Reparameterization Trick.png

To efficiently search for $\theta^*,\phi^* = \underset{\theta,\phi}\operatorname{arg max} \, L_{\theta,\phi}(x)$ the typical method is gradient ascent.

It is straightforward to find $\nabla_\theta \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{p_\theta(x, z)}{q_\phi({z| x})}\right]
= \mathbb E_{z \sim q_\phi(\cdot | x)} \left[ \nabla_\theta \ln \frac{p_\theta(x, z)}{q_\phi({z| x})}\right]$ However, $\nabla_\phi \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{p_\theta(x, z)}{q_\phi({z| x})}\right]$ does not allow one to put the $\nabla_\phi$ inside the expectation, since $\phi$ appears in the probability distribution itself. The reparameterization trick (also known as stochastic backpropagation{{Cite journal |last1=Rezende |first1=Danilo Jimenez |last2=Mohamed |first2=Shakir |last3=Wierstra |first3=Daan |date=2014-06-18 |title=Stochastic Backpropagation and Approximate Inference in Deep Generative Models |url=https://proceedings.mlr.press/v32/rezende14.html |journal=International Conference on Machine Learning |language=en |publisher=PMLR |pages=1278–1286|arxiv=1401.4082 }}) bypasses this difficulty.{{Cite journal|last1=Bengio|first1=Yoshua|last2=Courville|first2=Aaron|last3=Vincent|first3=Pascal|title=Representation Learning: A Review and New Perspectives|url=https://ieeexplore.ieee.org/document/6472238|journal=IEEE Transactions on Pattern Analysis and Machine Intelligence|year=2013|volume=35|issue=8|pages=1798–1828|doi=10.1109/TPAMI.2013.50|pmid=23787338|issn=1939-3539|arxiv=1206.5538|s2cid=393948}}{{Cite arXiv|last1=Kingma|first1=Diederik P.|last2=Rezende|first2=Danilo J.|last3=Mohamed|first3=Shakir|last4=Welling|first4=Max|date=2014-10-31|title=Semi-Supervised Learning with Deep Generative Models|class=cs.LG|eprint=1406.5298}}

The most important example is when $z \sim q_\phi(\cdot | x)$ is normally distributed, as $\mathcal N(\mu_\phi(x), \Sigma_\phi(x))$ .

: File:Reparameterized Variational Autoencoder.png

This can be reparametrized by letting $\boldsymbol{\varepsilon} \sim \mathcal{N}(0, \boldsymbol{I})$ be a "standard random number generator", and construct $z$ as $z = \mu_\phi(x) + L_\phi(x)\epsilon$ . Here, $L_\phi(x)$ is obtained by the Cholesky decomposition: $\Sigma_\phi(x) = L_\phi(x)L_\phi(x)^T$ Then we have $\nabla_\phi \mathbb E_{z \sim q_\phi(\cdot | x)} \left[\ln \frac{p_\theta(x, z)}{q_\phi({z| x})}\right]
=
\mathbb {E}_{\epsilon}\left[ \nabla_\phi \ln {\frac {p_{\theta }(x, \mu_\phi(x) + L_\phi(x)\epsilon)}{q_{\phi }(\mu_\phi(x) + L_\phi(x)\epsilon | x)}}\right]$ and so we obtained an unbiased estimator of the gradient, allowing stochastic gradient descent.

Since we reparametrized $z$ , we need to find $q_\phi(z|x)$ . Let $q_0$ be the probability density function for $\epsilon$ , then {{clarify |reason=The following calculations might have mistakes.|date=October 2023}} $\ln q_\phi(z | x) = \ln q_0 (\epsilon) - \ln|\det(\partial_\epsilon z)|$ where $\partial_\epsilon z$ is the Jacobian matrix of $z$ with respect to $\epsilon$ . Since $z = \mu_\phi(x) + L_\phi(x)\epsilon$ , this is $\ln q_\phi(z | x) = -\frac 12 \|\epsilon\|^2 - \ln|\det L_\phi(x)| - \frac n2 \ln(2\pi)$

Variations

Many variational autoencoders applications and extensions have been used to adapt the architecture to other domains and improve its performance.

$\beta$ -VAE is an implementation with a weighted Kullback–Leibler divergence term to automatically discover and interpret factorised latent representations. With this implementation, it is possible to force manifold disentanglement for $\beta$ values greater than one. This architecture can discover disentangled latent factors without supervision.{{Cite conference|last1=Higgins|first1=Irina|last2=Matthey|first2=Loic|last3=Pal|first3=Arka|last4=Burgess|first4=Christopher|last5=Glorot|first5=Xavier|last6=Botvinick|first6=Matthew|last7=Mohamed|first7=Shakir|last8=Lerchner|first8=Alexander|date=2016-11-04|title=beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework|url=https://openreview.net/forum?id=Sy2fzU9gl|language=en|conference=NeurIPS}}{{Cite arXiv|last1=Burgess|first1=Christopher P.|last2=Higgins|first2=Irina|last3=Pal|first3=Arka|last4=Matthey|first4=Loic|last5=Watters|first5=Nick|last6=Desjardins|first6=Guillaume|last7=Lerchner|first7=Alexander|date=2018-04-10|title=Understanding disentangling in β-VAE|class=stat.ML|eprint=1804.03599}}

The conditional VAE (CVAE), inserts label information in the latent space to force a deterministic constrained representation of the learned data.{{Cite conference|last1=Sohn|first1=Kihyuk|last2=Lee|first2=Honglak|last3=Yan|first3=Xinchen|date=2015-01-01|title=Learning Structured Output Representation using Deep Conditional Generative Models|url=https://proceedings.neurips.cc/paper/2015/file/8d55a249e6baa5c06772297520da2051-Paper.pdf|language=en|conference=NeurIPS}}

Some structures directly deal with the quality of the generated samples{{Cite arXiv|last1=Dai|first1=Bin|last2=Wipf|first2=David|date=2019-10-30|title=Diagnosing and Enhancing VAE Models|class=cs.LG|eprint=1903.05789}}{{Cite arXiv|last1=Dorta|first1=Garoe|last2=Vicente|first2=Sara|last3=Agapito|first3=Lourdes|last4=Campbell|first4=Neill D. F.|last5=Simpson|first5=Ivor|date=2018-07-31|title=Training VAEs Under Structured Residuals|class=stat.ML|eprint=1804.01050}} or implement more than one latent space to further improve the representation learning.

Some architectures mix VAE and generative adversarial networks to obtain hybrid models.{{Cite journal|last1=Larsen|first1=Anders Boesen Lindbo|last2=Sønderby|first2=Søren Kaae|last3=Larochelle|first3=Hugo|last4=Winther|first4=Ole|date=2016-06-11|title=Autoencoding beyond pixels using a learned similarity metric|url=http://proceedings.mlr.press/v48/larsen16.html|journal=International Conference on Machine Learning|language=en|publisher=PMLR|pages=1558–1566|arxiv=1512.09300}}{{cite arXiv|last1=Bao|first1=Jianmin|last2=Chen|first2=Dong|last3=Wen|first3=Fang|last4=Li|first4=Houqiang|last5=Hua|first5=Gang|date=2017|title=CVAE-GAN: Fine-Grained Image Generation Through Asymmetric Training|pages=2745–2754|class=cs.CV|eprint=1703.10155}}{{Cite journal|last1=Gao|first1=Rui|last2=Hou|first2=Xingsong|last3=Qin|first3=Jie|last4=Chen|first4=Jiaxin|last5=Liu|first5=Li|last6=Zhu|first6=Fan|last7=Zhang|first7=Zhao|last8=Shao|first8=Ling|date=2020|title=Zero-VAE-GAN: Generating Unseen Features for Generalized and Transductive Zero-Shot Learning|url=https://ieeexplore.ieee.org/document/8957359|journal=IEEE Transactions on Image Processing|volume=29|pages=3665–3680|doi=10.1109/TIP.2020.2964429|pmid=31940538|bibcode=2020ITIP...29.3665G|s2cid=210334032|issn=1941-0042|url-access=subscription}}

It is not necessary to use gradients to update the encoder. In fact, the encoder is not necessary for the generative model. {{cite book | last1=Drefs | first1=J. | last2=Guiraud | first2=E. | last3=Panagiotou | first3=F. | last4=Lücke | first4=J. | chapter=Direct evolutionary optimization of variational autoencoders with binary latents | title=Joint European Conference on Machine Learning and Knowledge Discovery in Databases | series=Lecture Notes in Computer Science | pages=357–372 | year=2023 | volume=13715 | publisher=Springer Nature Switzerland | doi=10.1007/978-3-031-26409-2_22 | isbn=978-3-031-26408-5 | chapter-url=https://link.springer.com/chapter/10.1007/978-3-031-26409-2_22 }}

Statistical distance VAE variants

After the initial work of Diederik P. Kingma and Max Welling,{{Cite arXiv |eprint=1312.6114 |class=stat.ML |first1=Diederik P. |last1=Kingma |first2=Max |last2=Welling |title=Auto-Encoding Variational Bayes |date=2022-12-10}} several procedures were

proposed to formulate in a more abstract way the operation of the VAE. In these approaches the loss function is composed of two parts :

the usual reconstruction error part which seeks to ensure that the encoder-then-decoder mapping $x \mapsto D_\theta(E_\psi(x))$ is as close to the identity map as possible; the sampling is done at run time from the empirical distribution $\mathbb{P}^{real}$ of objects available (e.g., for MNIST or IMAGENET this will be the empirical probability law of all images in the dataset). This gives the term: $\mathbb{E}_{x \sim \mathbb{P}^{real}} \left[ \|x - D_\theta(E_\phi(x))\|_2^2\right]$ .
a variational part that ensures that, when the empirical distribution $\mathbb{P}^{real}$ is passed through the encoder $E_\phi$ , we recover the target distribution, denoted here $\mu(dz)$ that is usually taken to be a Multivariate normal distribution. We will denote $E_\phi \sharp \mathbb{P}^{real}$ this pushforward measure which in practice is just the empirical distribution obtained by passing all dataset objects through the encoder $E_\phi$ . In order to make sure that $E_\phi \sharp \mathbb{P}^{real}$ is close to the target $\mu(dz)$ , a Statistical distance $d$ is invoked and the term $d \left( \mu(dz), E_\phi \sharp \mathbb{P}^{real} \right)^2$ is added to the loss.

We obtain the final formula for the loss:

$L_{\theta,\phi} = \mathbb{E}_{x \sim \mathbb{P}^{real}} \left[ \|x - D_\theta(E_\phi(x))\|_2^2\right]
+d \left( \mu(dz), E_\phi \sharp \mathbb{P}^{real} \right)^2$

The statistical distance $d$ requires special properties, for instance it has to be posses a formula as expectation because the loss function will need to be optimized by stochastic optimization algorithms. Several distances can be chosen and this gave rise to several flavors of VAEs:

the sliced Wasserstein distance used by S Kolouri, et al. in their VAE{{Cite conference |last1=Kolouri |first1=Soheil |last2=Pope |first2=Phillip E. |last3=Martin |first3=Charles E. |last4=Rohde |first4=Gustavo K. |date=2019 |title=Sliced Wasserstein Auto-Encoders |url=https://openreview.net/forum?id=H1xaJn05FQ |conference=International Conference on Learning Representations |publisher=ICPR |book-title=International Conference on Learning Representations}}
the energy distance implemented in the Radon Sobolev Variational Auto-Encoder{{Cite journal |last=Turinici |first=Gabriel |year=2021 |title=Radon-Sobolev Variational Auto-Encoders |url=https://www.sciencedirect.com/science/article/pii/S0893608021001556 |journal=Neural Networks |volume=141 |pages=294–305 |arxiv=1911.13135 |doi=10.1016/j.neunet.2021.04.018 |issn=0893-6080 |pmid=33933889}}
the Maximum Mean Discrepancy distance used in the MMD-VAE{{Cite journal |arxiv=1705.02239 |first1=A. |last1=Gretton |first2=Y. |last2=Li |title=A Polya Contagion Model for Networks |date=2017 |last3=Swersky |first3=K. |last4=Zemel |first4=R. |last5=Turner |first5=R.|journal=IEEE Transactions on Control of Network Systems |volume=5 |issue=4 |pages=1998–2010 |doi=10.1109/TCNS.2017.2781467 }}
the Wasserstein distance used in the WAEs{{Cite arXiv |eprint=1711.01558 |first1=I. |last1=Tolstikhin |first2=O. |last2=Bousquet |title=Wasserstein Auto-Encoders |date=2018 |last3=Gelly |first3=S. |last4=Schölkopf |first4=B.|class=stat.ML }}
kernel-based distances used in the Kernelized Variational Autoencoder (K-VAE){{Cite arXiv |eprint=1901.02401 |first1=C. |last1=Louizos |first2=X. |last2=Shi |title=Kernelized Variational Autoencoders |date=2019 |last3=Swersky |first3=K. |last4=Li |first4=Y. |last5=Welling |first5=M.|class=astro-ph.CO }}