Adversarial Autoencoders

Alireza MakhzaniJonathon ShlensN. JaitlyI. Goodfellow

article2015arXiv2,400 citations

Proposes the adversarial autoencoder framework, which uses generative adversarial training to perform variational inference with arbitrary latent priors, enabling flexible generative modeling, semi-supervised classification, and style-content disentanglement.

Listen

Capturing complex real-world data patterns such as images, audio, and video requires scalable generative machine learning models. Traditional approaches often rely on slow statistical sampling methods that degrade during training, while modern alternatives can impose restrictive mathematical constraints or prove difficult to train. Building flexible systems that learn meaningful representations and perform well with limited human supervision remains a major computational challenge.

The article introduces and evaluates the adversarial autoencoder, a neural network architecture designed to turn standard autoencoders into generative models by using an adversarial training technique to match internal representations to target distributions.

The authors conducted empirical evaluations across standard benchmark image datasets, including MNIST handwritten digits, the Toronto Face dataset, and Street View House Numbers. The framework integrates an autoencoder, which compresses and reconstructs data, with an adversarial network that acts as a regularizer to ensure internal data encodings match chosen distributions without requiring explicit mathematical formulas. The approach was tested across generative modeling, semi-supervised classification, unsupervised clustering, and dimensionality reduction.

The article demonstrates several key findings. First, the proposed model outperformed prior generative methods in generative sample quality, achieving higher test log-likelihood scores on both MNIST and the Toronto Face dataset. Second, in semi-supervised classification with only 100 to 1,000 labeled examples, the architecture significantly improved upon standard variational autoencoders and matched or closely approached leading specialized techniques, achieving an error rate of 1.90% on MNIST with 100 labels. Third, the model successfully separated discrete content, such as digit class, from continuous style factors, such as writing angle. Fourth, in fully unsupervised clustering, the model achieved a 4.10% classification error rate when grouping data into 30 clusters. Finally, the framework produced smooth, continuous low-dimensional data embeddings that prevent the representational fractures typical of standard autoencoders.

These results show that organizations can achieve strong generative modeling and classification performance while drastically reducing data annotation costs. Unlike alternative frameworks that require layer-by-layer training or exact analytical probability formulas, this architecture trains end-to-end and accommodates arbitrary target distributions. This capability lowers software pipeline complexity, reduces labeling overhead, and improves data visualization for complex datasets.

Stakeholders and engineering teams seeking to implement this approach should prioritize deploying adversarial autoencoders for semi-supervised tasks where labeled data is scarce and expensive. To maintain stability, practitioners should apply batch normalization when handling complex, natural datasets. Before production deployment, further evaluation on broader domain-specific datasets is advisable to ensure robustness across diverse data distributions.

A primary limitation noted in the article is that standard statistical metrics for evaluating generative model likelihoods, such as Parzen window estimates, remain imperfect approximations. Additionally, when using deterministic encoders, the internal representation size must be carefully matched to the intrinsic dimensionality of the dataset. Despite these evaluation caveats, the qualitative and quantitative evidence provides high confidence in the architecture's effectiveness for semi-supervised learning, clustering, and data representation.

Google OpenAI University of Toronto classification data visualization dimensionality reduction Adversarial Distribution Autoencoder

arXiv: 1511.05644

Paper: Generative Adversarial Networks, Ian J. Goodfellow et al. (2014). Reading the foundational Generative Adversarial Networks paper first is essential because AAEs directly build upon its adversarial training framework to perform variational inference.
Paper: An Introduction to Variational Autoencoders, Diederik P. Kingma et al. (2019). Understanding variational autoencoders provides the necessary theoretical background on aggregated posteriors and latent variable optimization that adversarial autoencoders extend.
Paper: Conditional Generative Adversarial Nets, Mehdi Mirza et al. (2014). Conditional generative adversarial networks introduce the architectural conditioning techniques that AAEs adapt for semi-supervised classification and style disentanglement.

Paper: InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets, Xi Chen et al. (2016). InfoGAN directly continues the pursuit of unsupervised disentangled representation learning initiated by adversarial autoencoders, maximizing mutual information for interpretable latent codes.
Paper: beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework, Irina Higgins et al. (2016). This work extends the exploration of disentangled latent representations in probabilistic autoencoders by introducing a constrained variational framework with targeted regularization.
Paper: Isolating Sources of Disentanglement in Variational Autoencoders, Ricky T. Q. Chen et al. (2018). This paper generalizes and refines the disentanglement objectives of earlier latent variable models by decomposing the evidence lower bound to explicitly isolate total correlation.

Abstract

In this paper, we propose the "adversarial autoencoder" (AAE), which is a probabilistic autoencoder that uses the recently proposed generative adversarial networks (GAN) to perform variational inference by matching the aggregated posterior of the hidden code vector of the autoencoder with an arbitrary prior distribution. Matching the aggregated posterior to the prior ensures that generating from any part of prior space results in meaningful samples. As a result, the decoder of the adversarial autoencoder learns a deep generative model that maps the imposed prior to the data distribution. We show how the adversarial autoencoder can be used in applications such as semi-supervised classification, disentangling style and content of images, unsupervised clustering, dimensionality reduction and data visualization. We performed experiments on MNIST, Street View House Numbers and Toronto Face datasets and show that adversarial autoencoders achieve competitive results in generative modeling and semi-supervised classification tasks.

1 Introduction
1.1 Generative Adversarial Networks
2 Adversarial Autoencoders
2.1 Relationship to Variational Autoencoders
2.2 Relationship to GANs and GMMNs
2.3 Incorporating Label Information in the Adversarial Regularization
3 Likelihood Analysis of Adversarial Autoencoders
4 Supervised Adversarial Autoencoders
5 Semi-Supervised Adversarial Autoencoders
6 Unsupervised Clustering with Adversarial Autoencoders
7 Dimensionality Reduction with Adversarial Autoencoders
8 Conclusion
References
A Experiment Details
A.1 Likelihood Experiments
A.2 Semi-Supervised Experiments
A.2.1 MNIST
A.2.2 SVHN
A.3 Unsupervised Clustering Experiments

Knowls

Knowl 1 — Adversarial Autoencoder Framework
model/method
The Adversarial Autoencoder (AAE) is a generative autoencoder that regularizes the latent representation of an autoencoder by matching the aggregated posterior distribution of its latent code to an arbitrary user-defined prior distribution $p(z)$ via generative adversarial training.

Let $x \in \mathcal{X}$ denote the input data with data distribution $p_d(x)$ , and $z \in \mathbb{R}^d$ denote the latent code vector. The autoencoder consists of an encoder $q(z|x)$ and a decoder $p(x|z)$ . The encoder defines an aggregated posterior distribution $q(z)$ over the latent space:

$q(z) = \int_{\mathcal{X}} q(z|x) p_d(x) \, dx$

To regularize $q(z)$ to match the prior $p(z)$ , an adversarial discriminator $D(z)$ is attached to the latent representation. The encoder acts as the generator $G$ in the min-max adversarial objective:

$\min_G \max_D \mathbb{E}_{z \sim p(z)}[\log D(z)] + \mathbb{E}_{x \sim p_d(x)}[\log(1 - D(q(z|x)))]$

Training proceeds on each mini-batch in two alternating phases using stochastic gradient descent:
1. Reconstruction Phase: The encoder $q(z|x)$ and decoder $p(x|z)$ are updated to minimize the reconstruction error between the input $x$ and the reconstructed output $\tilde{x}$ .
2. Regularization Phase: The discriminator $D(z)$ is updated to distinguish true samples drawn from $p(z)$ from latent representations generated by the encoder $q(z|x)$ . The encoder is then updated to confuse the discriminator into classifying its output as samples from $p(z)$ .
After training, the decoder $p(x|z)$ functions as a generative model by passing random samples drawn from $p(z)$ through the decoder to map them to data space.
Knowl 2 — Posterior Encoder Formulations in Adversarial Autoencoders
model/method
In an Adversarial Autoencoder, the encoder $q(z|x)$ mapping an input $x$ to a latent code $z$ can be configured using three distinct formulations of conditional stochasticity:
1. Deterministic Encoder: $q(z|x)$ is a deterministic parameterized function $z = f(x)$ . The only source of stochasticity in the aggregated posterior $q(z) = \int q(z|x) p_d(x) dx$ is the empirical data distribution $p_d(x)$ . In this setting, the latent dimensionality must closely match the intrinsic data dimensionality.
2. Gaussian Posterior Encoder: $q(z|x) = \mathcal{N}(z; \mu(x), \Sigma(x))$ , where the encoder network outputs mean vectors $\mu(x)$ and diagonal variances $\Sigma(x) = \operatorname{diag}(\sigma_1^2(x), \dots, \sigma_d^2(x))$ . Stochasticity originates from both the data distribution and Gaussian sampling via the reparameterization trick $z = \mu(x) + \sigma(x) \odot \epsilon$ , where $\epsilon \sim \mathcal{N}(0, I)$ .
3. Universal Approximator Posterior Encoder: $q(z|x)$ is parameterized as a function $f(x, \eta)$ that accepts the data input $x$ and an independent noise vector $\eta \sim p_\eta(\eta)$ drawn from a fixed distribution (such as a standard Gaussian):
$q(z|x) = \int \delta(z - f(x, \eta)) p_\eta(\eta) \, d\eta \implies q(z) = \int_{\mathcal{X}} \int \delta(z - f(x, \eta)) p_d(x) p_\eta(\eta) \, d\eta \, dx$

This parameterization allows the encoder to learn an arbitrary, non-Gaussian posterior distribution $q(z|x)$ for each input $x$ , and the adversarial network trains the encoder via direct back-propagation through $f(x, \eta)$ .
Knowl 3 — Variational Autoencoder Decomposition and Relation to Adversarial Autoencoders
theoretical result
In Variational Autoencoders (VAEs), the negative log-likelihood of data $x \sim p_d(x)$ is upper-bounded by:

$\mathbb{E}_{x \sim p_d(x)}[-\log p(x)] \le \mathbb{E}_{x \sim p_d(x)}\left[\mathbb{E}_{q(z|x)}[-\log p(x|z)]\right] + \mathbb{E}_{x \sim p_d(x)}\left[\text{KL}(q(z|x) \parallel p(z))\right]$

Assuming a Gaussian conditional posterior $q(z|x) = \mathcal{N}(\mu(x), \operatorname{diag}(\sigma_1^2(x), \dots, \sigma_d^2(x)))$ and prior $p(z)$ , this variational bound decomposes into three components:

$\mathbb{E}_{x \sim p_d(x)}[-\log p(x)] \le \mathbb{E}_{x \sim p_d(x)}\left[\mathbb{E}_{q(z|x)}[-\log p(x|z)]\right] - \mathbb{E}_{x \sim p_d(x)}\left[\sum_{i=1}^d \log \sigma_i(x)\right] + \mathbb{E}_{q(z)}[-\log p(z)] + \text{const}$

$= \text{Reconstruction Error} - \text{Entropy}(q(z|x)) + \text{CrossEntropy}(q(z), p(z))$

where $q(z) = \int q(z|x) p_d(x) dx$ is the aggregated posterior distribution.

In VAEs, the explicit KL divergence / cross-entropy term encourages $q(z)$ to match the modes of $p(z)$ and requires an exact analytic functional form for $p(z)$ . Adversarial Autoencoders (AAEs) replace the entropy and cross-entropy penalty terms with an adversarial min-max objective that matches the aggregated posterior $q(z)$ to the complete distribution of $p(z)$ . This enables the imposition of arbitrary priors (including non-parametric distributions) that can only be sampled from, preventing unpopulated regions ("holes") in the latent coding space.
Knowl 4 — Semi-Supervised Adversarial Autoencoders
model/method
The Semi-Supervised Adversarial Autoencoder (Semi-Supervised AAE) utilizes unlabeled data in conjunction with a small proportion of labeled data to learn disentangled representations of categorical class identity and continuous style variation.

The generative model assumes that observed data $x$ is generated by a discrete class variable $y \sim \text{Cat}(y)$ and an independent continuous latent variable $z \sim \mathcal{N}(0, I)$ . The encoder network $q(y, z | x)$ outputs both a categorical distribution over labels via a softmax layer and a continuous style vector $z$ via a linear layer. The decoder $p(x | y, z)$ reconstructs the input from the combination of the one-hot label vector and the continuous style vector $z$ .

Regularization is enforced by two separate adversarial discriminators:
1. Categorical Adversarial Network: Matches the aggregated posterior distribution $q(y) = \int q(y|x) p_d(x) dx$ to the Categorical prior distribution $\text{Cat}(y)$ , ensuring $y$ does not encode continuous style information.
2. Gaussian Adversarial Network: Matches the aggregated posterior $q(z) = \int q(z|x) p_d(x) dx$ to the standard Gaussian prior $\mathcal{N}(0, I)$ .
Training alternates through three phases on each mini-batch:
1. Reconstruction Phase: On an unlabeled mini-batch, the encoder $q(y, z|x)$ and decoder $p(x|y, z)$ are updated to minimize the reconstruction error.
2. Regularization Phase: The two discriminators are updated to classify true prior samples from generated latent vectors $y$ and $z$ , and the encoder is updated to confuse both discriminators.
3. Semi-Supervised Classification Phase: On a labeled mini-batch $(x, y_{\text{true}})$ , the encoder $q(y|x)$ is updated to minimize the cross-entropy classification error against ground-truth labels $y_{\text{true}}$ .
Knowl 5 — Supervised Style and Content Disentanglement in Adversarial Autoencoders
model/method
In supervised Adversarial Autoencoders, label information is disentangled from latent style by routing the ground-truth one-hot label vector directly to the decoder network alongside the continuous latent code $z = q(z|x)$ . The encoder $q(z|x)$ is regularized via an adversarial discriminator to match an imposed Gaussian prior (e.g., $\mathcal{N}(0, I)$ ).

Because the decoder receives the label information explicitly, the encoder is constrained to retain only label-independent factors of variation in $z$ , such as stroke thickness, slant, and writing style in images of digits. Once trained, setting the latent code $z$ to a fixed value while systematically sweeping the one-hot label vector input generates synthetic samples across all target classes that preserve the identical style properties.
Knowl 6 — Label-Conditioned Latent Distribution Regularization
model/method
Adversarial Autoencoders can map distinct data classes to specific components or geometric modes of an arbitrary prior distribution (such as a mixture of 10 Gaussians or a continuous Swiss roll manifold) by providing label information to the adversarial discriminator.

A one-hot conditioning vector is appended to the input of the discriminator. When the prior contains $K$ mixture components, the one-hot vector has dimension $K+1$ , where the $(K+1)$ -th entry designates unlabeled samples:
1. Positive Phase (True Prior Samples): For labeled examples of class $i$ , true samples are drawn from the $i$ -th prior component and the discriminator is provided the one-hot vector for class $i$ . For unlabeled examples, true samples are drawn from the full prior mixture and the $(K+1)$ -th dimension is activated.
2. Negative Phase (Generated Codes): The latent code $z \sim q(z|x)$ produced by the encoder is paired with the corresponding one-hot class label (for labeled inputs) or the $(K+1)$ -th unlabeled indicator (for unlabeled inputs) and fed to the discriminator.
This conditioning forces each data class to occupy its dedicated mode or manifold region in latent space while maintaining consistent style geometry across classes.
Knowl 7 — Unsupervised Clustering with Adversarial Autoencoders
model/method
Unsupervised clustering in Adversarial Autoencoders disentangles discrete cluster assignments from continuous style representations without requiring any label supervision.

The architecture uses an inference encoder $q(y, z|x)$ where $y \in \{0, 1\}^K$ is a discrete $K$ -dimensional one-hot vector output by a softmax layer (representing $K$ cluster assignments), and $z \in \mathbb{R}^d$ is a continuous latent style vector. The decoder reconstructs $x$ conditioned on both $y$ and $z$ . The latent space is regularized by two adversarial discriminators:
1. A categorical discriminator that matches the aggregated posterior $q(y) = \int q(y|x) p_d(x) dx$ to a uniform Categorical prior distribution $\text{Cat}(y)$ .
2. A Gaussian discriminator that matches the aggregated posterior $q(z) = \int q(z|x) p_d(x) dx$ to $\mathcal{N}(0, I)$ .
Training alternates between the reconstruction phase and the regularization phase across unlabeled data. No semi-supervised cross-entropy phase is used.

To assign class names to learned clusters for evaluation, for each cluster $i \in \{1, \dots, K\}$ , the validation image $x_n$ that maximizes $q(y_i|x_n)$ is identified, and the ground-truth label of $x_n$ is assigned to all points in cluster $i$ . Test error is then evaluated on the hold-out dataset according to the assigned cluster identities.
Knowl 8 — Dimensionality Reduction and Manifold Regularization with Adversarial Autoencoders
model/method
Adversarial Autoencoders eliminate the problem of manifold fracturing (where non-regularized autoencoders separate continuous data manifolds into disjoint domains) by structuring the latent representation into an additive combination of learned cluster centroids and continuous style vectors.

For a dataset with $m$ categories mapped to an $n$ -dimensional latent space:
1. The encoder predicts a discrete class probability vector $y \in [0, 1]^m$ and an $n$ -dimensional style vector $z \in \mathbb{R}^n$ .
2. An $m \times n$ matrix $W_C$ defines $m$ cluster head vectors in $\mathbb{R}^n$ . The categorical output $y$ selects a cluster head representation via $y^T W_C$ .
3. The final latent representation is defined additively as $z_{\text{final}} = y^T W_C + z$ .
4. Two adversarial discriminators enforce a Categorical distribution on $y$ and a Gaussian distribution $\mathcal{N}(0, I)$ on $z$ .
5. An explicit separation cost is applied to $W_C$ : for every pair of cluster heads with Euclidean distance $d_{ij}$ , the cost is $\max(0, \eta - d_{ij})$ , penalizing pairwise distances that are smaller than a threshold $\eta$ .
This configuration ensures that cluster centers are separated by at least distance $\eta$ while the local distribution around each cluster center forms a continuous Gaussian manifold.

Knowl 9 — Generative Log-Likelihood Evaluation on MNIST and TFD

data/table

The generative modeling capability of the Adversarial Autoencoder (AAE) was evaluated by computing test data log-likelihood on real-valued MNIST and the Toronto Face Dataset (TFD). A Gaussian Parzen window (kernel density estimator) was fit to 10,000 (10K) and 10,000,000 (10M) samples generated by the trained models, with the Parzen window bandwidth $\sigma$ selected by cross-validation.

Model	MNIST (10K)	MNIST (10M)	TFD (10K)	TFD (10M)
DBN	$138 \pm 2$	-	$1909 \pm 66$	-
Stacked CAE	$121 \pm 1.6$	-	$2110 \pm 50$	-
Deep GSN	$214 \pm 1.1$	-	$1890 \pm 29$	-
GAN	$225 \pm 2$	$386$	$2057 \pm 26$	-
GMMN + AE	$282 \pm 2$	-	$2204 \pm 20$	-
Adversarial Autoencoder	$\mathbf{340 \pm 2}$	$\mathbf{427}$	$\mathbf{2252 \pm 16}$	$\mathbf{2522}$

Higher log-likelihood values indicate closer approximation of the data distribution. The AAE outperformed Deep Belief Networks (DBN), Stacked Contractive Autoencoders (Stacked CAE), Deep Generative Stochastic Networks (Deep GSN), standard GANs, and Generative Moment Matching Networks with Autoencoders (GMMN + AE) across both datasets and sample evaluation budgets.

Knowl 10 — Semi-Supervised Classification and Clustering Performance Benchmarks

data/table

Semi-supervised classification error rates were measured on MNIST (with 100, 1000, and all labeled training examples) and Street View House Numbers (SVHN, with 1000 labeled examples). Unsupervised clustering error on MNIST was evaluated across different cluster numbers.

Method	MNIST (100)	MNIST (1000)	MNIST (All)	SVHN (1000)
NN Baseline	25.80%	8.73%	1.25%	47.50%
VAE (M1) + TSVM	$11.82 \pm 0.25$ %	$4.24 \pm 0.07$ %	-	$55.33 \pm 0.11$ %
VAE (M2)	$11.97 \pm 1.71$ %	$3.60 \pm 0.56$ %	-	-
VAE (M1 + M2)	$3.33 \pm 0.14$ %	$2.40 \pm 0.02$ %	0.96%	$36.02 \pm 0.10$ %
VAT	2.33%	1.36%	$0.64 \pm 0.04$ %	24.63%
CatGAN	$1.91 \pm 0.1$ %	$1.73 \pm 0.18$ %	0.91%	-
Ladder Networks	$1.06 \pm 0.37$ %	$0.84 \pm 0.08$ %	$0.57 \pm 0.02$ %	-
ADGM	$0.96 \pm 0.02$ %	-	-	$\mathbf{16.61 \pm 0.24}$ %
Adversarial Autoencoders	$1.90 \pm 0.10$ %	$1.60 \pm 0.08$ %	$0.85 \pm 0.02$ %	$17.70 \pm 0.30$ %

For purely unsupervised clustering on MNIST, AAE achieved an error rate of $9.55 \pm 2.05$ % with 16 clusters and $4.10 \pm 1.13$ % with 30 clusters (compared to 9.70% for CatGAN with 20 clusters). On SVHN with 1000 labels, AAE achieved 17.70% error, matching the performance of the Auxiliary Deep Generative Model (ADGM at 16.61%) while training end-to-end via back-propagation rather than layer-wise.

Coverage note — No substantial contributed material was omitted.

References

1.Geoffrey E. Hinton, Simon Osindero, and Yee Whye Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18:1527–1554, 2006.
2.Ruslan Salakhutdinov and Geoffrey E Hinton. Deep boltzmann machines. In International Conference on Artificial Intelligence and Statistics, pages 448–455, 2009.
3.Diederik P Kingma and Max Welling. Auto-encoding variational bayes. International Conference on Learning Representations (ICLR), 2014.
4.Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. International Conference on Machine Learning, 2014.
5.Yuri Burda, Roger Grosse, and Ruslan Salakhutdinov. Importance weighted autoencoders. arXiv preprint arXiv:1509.00519, 2015.
6.Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pages 2672–2680, 2014.
7.Yujia Li, Kevin Swersky, and Richard Zemel. Generative moment matching networks. International Conference on Machine Learning (ICML), 2015.
8.Yoshua Bengio, Grégoire Mesnil, Yann Dauphin, and Salah Rifai. Better mixing via deep representations. International Conference on Machine Learning (ICML), 2013.
9.Yoshua Bengio, Eric Thibodeau-Laufer, Guillaume Alain, and Jason Yosinski. Deep generative stochastic networks trainable by backprop. International Conference on Machine Learning (ICML), 2014.
10.Lucas Theis, Aäron van den Oord, and Matthias Bethge. A note on the evaluation of generative models. arXiv preprint arXiv:1511.01844, 2015.
11.Diederik P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. In Advances in Neural Information Processing Systems, pages 3581–3589, 2014.
12.Lars Maaløe, Casper Kaae Sønderby, Søren Kaae Sønderby, and Ole Winther. Auxiliary deep generative models. arXiv preprint arXiv:1602.05473, 2016.
13.Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning, volume 2011, page 5. Granada, Spain, 2011.
14.Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, Ken Nakae, and Shin Ishii. Distributional smoothing with virtual adversarial training. stat, 1050:25, 2015.
15.Jost Tobias Springenberg. Unsupervised and semi-supervised learning with categorical generative adversarial networks. arXiv preprint arXiv:1511.06390, 2015.
16.Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semi-supervised learning with ladder networks. In Advances in Neural Information Processing Systems, pages 3532–3540, 2015.
17.Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(2579-2605):85, 2008.
18.Laurens Maaten. Learning a parametric embedding by preserving local structure. In International Conference on Artificial Intelligence and Statistics, pages 384–391, 2009.
19.Geoffrey Hinton. Non-linear dimensionality reduction. https://www.cs.toronto.edu/ hinton/csc2535/notes/lec11new.pdf.
20.Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from tensorflow.org.
21.Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.

Citation

MLA

Makhzani, A., et al. “Adversarial Autoencoders”. arXiv, 2015, https://doi.org/10.48550/arxiv.1511.05644.

APA

Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., & Frey, B. (2015). Adversarial Autoencoders. arXiv. https://doi.org/10.48550/arxiv.1511.05644

Chicago

Makhzani, A., J. Shlens, N. Jaitly, I. Goodfellow, and B. Frey. 2015. “Adversarial Autoencoders”. Preprint, ArXiv. https://doi.org/10.48550/arxiv.1511.05644.

Harvard

Makhzani, A. et al. (2015) “Adversarial Autoencoders”. arXiv. Available at: https://doi.org/10.48550/arxiv.1511.05644.

Vancouver

1. Makhzani A, Shlens J, Jaitly N, Goodfellow I, Frey B (2015) Adversarial Autoencoders. https://doi.org/10.48550/arxiv.1511.05644

BibTeX

@misc{https://doi.org/10.48550/arxiv.1511.05644,
  doi = {10.48550/ARXIV.1511.05644},
  url = {https://arxiv.org/abs/1511.05644},
  author = {Makhzani, Alireza and Shlens, Jonathon and Jaitly, Navdeep and Goodfellow, Ian and Frey, Brendan},
  keywords = {Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Adversarial Autoencoders},
  publisher = {arXiv},
  year = {2015},
  copyright = {arXiv.org perpetual, non-exclusive license}
}

Metadata:DOI registry

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Published with permission

Adversarial Autoencoders

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Adversarial Autoencoder Framework

Knowl 2 — Posterior Encoder Formulations in Adversarial Autoencoders

Knowl 3 — Variational Autoencoder Decomposition and Relation to Adversarial Autoencoders

Knowl 4 — Semi-Supervised Adversarial Autoencoders

Knowl 5 — Supervised Style and Content Disentanglement in Adversarial Autoencoders

Knowl 6 — Label-Conditioned Latent Distribution Regularization