Joint-stochastic-approximation Autoencoders with Application to Semi-supervised Learning

Wenbo He; Zhijian Ou

REVIEW 3 major objections 5 minor 13 references

Joint-stochastic-approximation Autoencoders with Application to Semi-supervised Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A new autoencoder family trains deep generative models by directly maximizing data log-likelihood, and its discrete-latent version matches state-of-the-art semi-supervised classifiers on MNIST and SVHN.

desk verdict A useful but under-supported SSL extension of the JSA framework; the empirical claims rest on 10-step MCMC draws that are likely not mixing, and Algorithm 1 doesn't match the stated objective. read the letter →

arxiv 2505.18558 v1 pith:63MYTV44 submitted 2025-05-24 cs.LG stat.ML

Wenbo He , Zhijian Ou This is my paper

classification cs.LGstat.ML

keywords jointstochasticapproximationsemi-supervisedlearningdeepgenerativemodelsmaximumlikelihooddiscretelatentvariablesautoencodersMetropolisindependencesamplerinclusiveKLdivergence

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper formally introduces JSA autoencoders (JAEs), a family of deep directed generative models trained by joint stochastic approximation. Their training objective is the data log-likelihood itself, not a variational lower bound or an adversarial divergence, and the inference network is trained by minimizing the inclusive KL divergence between the posterior and the proposal. Because sampling is done with a Metropolis independence sampler whose proposal is the inference network, the algorithm handles discrete and continuous latent variables with the same machinery, without REINFORCE, Gumbel-softmax, or straight-through estimators. On semi-supervised MNIST with 100 labels and SVHN with 1000 labels, a JAE with a Bernoulli latent space reaches error rates around 1.96% and 6.22%, matching state-of-the-art DGMs that use continuous latent spaces. The paper claims this is the first successful application of discrete latent variable models to these challenging semi-supervised tasks.

What carries the argument

The engine is joint stochastic approximation (JSA). The learning problem is written as finding a root of a system of equations: the generative gradient $\mathbb{E}_{p_\theta(h|x)}[\frac{\partial}{\partial\theta}\log p_\theta(x,h)] = 0$ and the inference gradient $\mathbb{E}_{p_\theta(h|x)}[\frac{\partial}{\partial\phi}\log q_\phi(h|x)] = 0$. These are expectations under the posterior, so the Robbins-Monro stochastic approximation algorithm can solve them by sampling from a Markov kernel whose invariant distribution is $p(z;\lambda)$. The paper uses a Metropolis independence sampler with proposal $q_\phi(h|x)$ from the inference network; computing the importance ratio cancels the intractable marginal $p_\theta(x)$, and the accept/reject step makes the inference network a valid proposal even when it mismatches the true posterior.

What would settle it

Train the MNIST JAE while logging the fraction of proposed samples accepted by the Metropolis independence sampler and the per-unit flip rate in the 60-dimensional Bernoulli latent space. Low acceptance rates or no flips within the ten warm-up steps would show that the gradient estimates are not drawn from the posterior, contradicting the maximum-likelihood interpretation; high acceptance with fast mixing would support the claim.

Watch

Extended reading notes

Core claim

The central claim is that maximum-likelihood learning for a deep directed generative model can be cast as a stochastic approximation root-finding problem over paired generative and inference parameters, and that solving it yields an autoencoder which directly maximizes the data log-likelihood while simultaneously minimizing $\mathrm{KL}(p_\theta(h|x) \| q_\phi(h|x))$. The authors prove consistency in the nonparametric limit: when both models have infinite capacity and the data size grows, the generative model recovers the true data distribution and the inference model recovers the exact posterior. Empirically, they show that the Metropolis independence sampler compensates for encoder/decoder mismatch, that the same update rules work for discrete and continuous variables, and that a JAE with a 60-dimensional Bernoulli prior on MNIST and a 220-dimensional Bernoulli prior on SVHN matches the semi-supervised classification performance of continuous-latent DGMs. They interpret this as first evidence that discrete latent variable models can succeed in semi-supervised learning.

Load-bearing premise

The training procedure assumes that ten warm-up sampling steps per data point bring the sampler close enough to the posterior that the minibatch gradients approximate the true maximum-likelihood direction; if the chain is still far from equilibrium, the implemented algorithm is not actually maximizing the data log-likelihood.

Editorial extensions

If this is right

JAE training optimizes the data log-likelihood directly, so the fitted generative model inherits the statistical consistency properties of maximum likelihood instead of optimizing a surrogate bound.
Discrete latent variables become usable without gradient tricks such as REINFORCE, Gumbel-softmax, or straight-through estimators, making Bernoulli or categorical codes practical in semi-supervised learning.
Encoder/decoder structure mismatch does not bias the learned generative model, because the Metropolis independence sampler corrects the proposal distribution rather than being forced to match the posterior exactly.
Semi-supervised JAEs can disentangle class and style in the latent space and perform class-conditional generation with discrete codes, as shown by the MNIST traversal experiments.
The same learning criterion extends naturally to semi-supervised objectives by adding supervised likelihood terms and a discriminative classification term $\alpha \log q_\phi(y|x)$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

One testable extension is applying JAEs to text or categorical observations, where VAEs are known to suffer posterior collapse; the paper's context-free-grammar experiment hints that the MIS mechanism may avoid that failure mode.
A direct diagnostic that would strengthen or weaken the central claim is measuring the acceptance rate and mixing speed of the Metropolis independence sampler during training, which the paper does not report.
The nonparametric consistency argument suggests JAEs with continuous latent spaces should also improve over VAEs on likelihood-based evaluation; the binary MNIST log-likelihood comparison points in that direction but does not establish a clear advantage.
If the sampling assumption holds, the practical payoff is that latent-space choices can be driven by data structure rather than by the availability of gradient estimators, which would simplify model design for clustered, categorical, or symbolic data.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know before you pick this up. The core learning rule is not new—it is JSA from Xu & Ou (2016) and Wang et al. (2017), and the authors freely say so. The new pieces are the semi-supervised adaptation, the mixed discrete/continuous demonstrations, and the claim that Bernoulli-latent JAEs are competitive with continuous-latent DGMs on MNIST and SVHN. The second thing: the implemented algorithm probably does not do what the theory says. Algorithm 1 draws one MIS sample per data point after a 10-step warm-up in a 60- or 220-dimensional Bernoulli space. That is almost certainly not stationary, and no acceptance rates or chain diagnostics are reported. If those chains are biased, the θ update is effectively an expectation under qφ, not pθ, so the 'directly maximizes data log-likelihood' claim is unsupported.

The paper does earn credit. The factor-analysis experiment is a clean, convincing demonstration that JAE can recover the true generative model even when the encoder is a diagonal Gaussian and the posterior is correlated—something VAE cannot do under the same mismatch. The GMM experiment shows the mixed discrete/continuous prior behaves as intended, and the CFG experiment shows JAE generating syntactically coherent expressions where a Gumbel-softmax VAE collapses. Those are real within-paper results.

The soft spots are easy to name. Proposition 1 is just a recasting of Eq. (3) as a root-finding problem; Proposition 2 is the infinite-capacity consistency statement with no convergence analysis of the SA process. More worrying is the algorithm-objective gap. Eq. (4)-(5) define the SSL objective, but Algorithm 1's φ update for the supervised minibatch uses log qφ(y,h|x) rather than the log qφ(h|x,y) implied by the derivation, and the α weighting appears to double-count the classifier gradient. The SVHN number also depends on extra regularizers—a discriminative confidence loss and a pseudo-loss—that are not in the JSA objective. The MCMC mixing issue is the load-bearing one, and it is not addressed. The benchmark results are decent but not state-of-the-art; the MNIST 1.96% and SVHN 6.22% are well behind BadGAN's 0.80% and 4.25%, respectively.

Who should read it? Anyone working on discrete latent variable models or on stochastic-approximation training for deep generative models. The framework offers a genuine alternative to REINFORCE and Gumbel-softmax tricks, and the synthetic experiments are informative. But the paper needs a serious revision: code, MCMC diagnostics, a corrected Algorithm 1, and a comparison against discrete-latent SSL baselines like CatGAN. The 'first demonstration' claim also needs a tighter scope. As is, I would send it to review rather than desk reject—the core idea is legitimate and the empirical work, despite the gaps, is not a toy—but I would expect major revisions before it is publishable.

Referee Report

3 major / 5 minor

Summary. The paper introduces Joint-stochastic-approximation autoencoders (JAEs), a family of deep directed generative models trained by stochastic approximation. The JSA learning rule pairs maximum-likelihood estimation of the generative model with minimization of the inclusive KL divergence between the posterior and an inference model; the required posterior expectations are obtained by a Metropolis independence sampler using the inference model as proposal. The authors extend this framework to semi-supervised learning by adding a classification term and report experiments on synthetic factor analysis, a Gaussian mixture, discrete sequence data, and MNIST/SVHN semi-supervised benchmarks with Bernoulli latent spaces. The paper claims that JAEs with discrete latent variables achieve performance comparable to continuous-latent DGMs and that the method is robust to encoder/decoder structure mismatch.

Significance. If the central claim is valid, the paper offers a principled alternative to VAE and GAN training that is directly tied to the data log-likelihood, handles discrete latents without gradient estimators, and is robust to mismatch between the inference and generative models. The derivation in Section 3.2 is self-contained: the gradient equations are obtained from log-likelihood and inclusive KL, and the SA formulation is reasonable. The synthetic factor-analysis experiment is a nice controlled demonstration that MIS can compensate for a misspecified inference model. However, the main empirical claim on MNIST/SVHN rests on the assumption that the MIS chains reach stationarity within the 10-step warm-up used in Algorithm 1, and the paper provides no diagnostics to support this. Because that assumption is load-bearing for the claim that the implemented method maximizes the data log-likelihood, the significance of the empirical results is currently conditional on additional evidence.

major comments (3)

[Algorithm 1, Appendix B.4-B.5] The gradient updates in Algorithm 1 require samples from the posteriors pθ(y,h|x) and pθ(h|x,y). In the implementation described in Appendices B.4 and B.5, the MIS chain is restarted for every data point with a 10-step warm-up, using the factorized inference model qφ(y|x)qφ(h|x) as proposal. For a 60-dimensional Bernoulli space (MNIST) or 220-dimensional space (SVHN), ten component-wise Metropolis steps are very unlikely to have reached stationarity, particularly early in training when qφ is far from pθ. The paper reports no acceptance rates, effective sample sizes, or sensitivity to warm-up length. If the chains are biased, the θ update estimates E_qφ[∂θ log pθ(x,h)] rather than E_pθ[∂θ log pθ(x,h)], so the claim that the implemented algorithm 'directly maximizes the data log-likelihood' is not established. Please provide chain diagnostics and a warm-up sensitivity study, or substantially soften the claim.
[Section 3.2, Proposition 2] Proposition 2 is a nonparametric consistency statement that assumes infinite capacity and that the KL divergences can be driven to zero. It does not provide a convergence theorem for the SA recursion, nor does it address the finite-capacity, finite-sample regime used in the experiments. As stated, the proposition essentially restates the classical consistency of maximum likelihood in the infinite limit and does not quantify the bias introduced by finite MCMC sampling. The paper would be strengthened by a convergence argument for the SA update under verifiable conditions, or by an explicit statement that the theoretical guarantee is limited to the idealized nonparametric limit.
[Section 3.3, Eq. (4), and abstract] The abstract and Section 3.2 state that JSA learning directly maximizes the data log-likelihood, but the semi-supervised objective in Eq. (4) includes the term -α Σ_{(x,y)∈L} log qφ(y|x), and the SVHN experiments in Appendix B.5 add further regularizers (entropy minimization and a pseudo-loss). The θ-update in Algorithm 1 is indeed a likelihood gradient, so the core mechanism is intact, but the overall implemented objective is a regularized hybrid rather than the raw data log-likelihood. The paper should clarify this distinction in the abstract and Section 3.3 to avoid overstating the maximum-likelihood claim.

minor comments (5)

[Section 3.2] There is a typo in the phrase 'minimizes the inclusive KL divergence the between the posteriori and the inference model'; 'the between' should be removed and 'posteriori' should be 'posterior'.
[Section 4.1] The sentence 'the decoder qφ(h|x) is implemented as a 2d Gaussian' appears to refer to the inference network (encoder), not the decoder; please correct the terminology.
[Table 1 and Section 4.4] The asterisks in Table 1 are not defined, and the text says the SVHN error is around 6.80% on RGB images while the table lists 6.22±0.55 for JAE+Bernoulli; the relationship between the grayscale and RGB results and the meaning of the asterisks should be stated explicitly.
[Table 2 caption] The caption contains a typo: 'Evalution' should be 'Evaluation'.
[References] Several references have inconsistent formatting, e.g., 'arXiv: Artificial Intelligence' for Che et al. and 'arXiv: Machine Learning' for Goodfellow et al.; please unify the reference style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: JAE's likelihood and inclusive-KL derivation is re-derived in-paper, and benchmark results are external comparisons.

full rationale

The paper's central derivation is self-contained. Section 3.2 begins from the exact identity ∂/∂θ log pθ(x) = E_{pθ(h|x)}[∂/∂θ log pθ(x,h)], defines the joint optimization in Eq. (2), derives the root-finding system Eq. (3), and verifies it fits the stochastic-approximation form of Eq. (1). The Metropolis independence sampler is constructed with target p(z;λ) and proposal q(z;λ), so the accept/reject step is standard MCMC rather than an imported conclusion. Proposition 2 is a standard consistency statement under nonparametric limits, not a fitted prediction. The SSL extension in Section 3.3 similarly derives Eq. (5) from Eq. (4) and Algorithm 1. Self-citations to Xu & Ou (2016) and Wang et al. (2017) identify the origin of JSA and the multi-move SA technique, but the load-bearing equations are presented and proved in this paper, so the self-citation is not load-bearing. The empirical claims (Tables 1-2, Figures 1-5) are comparisons against external methods and datasets, not re-statements of fitted inputs. The 10-step MCMC warm-up noted in Appendices B.4-B.5 raises a question about whether the implemented sampler is near stationarity; that is an empirical correctness risk, not definitional circularity, and cannot by itself support a circularity finding under the requirement to exhibit a specific reduction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central derivation relies on standard stochastic-approximation regularity conditions, on the MCMC sampler being at stationarity, and on q covering the posterior. In practice the paper also introduces at least four hand-chosen numbers: alpha, the SVHN regularization weights, the inference-noise variance schedule, and the MIS warm-up length, none of which are part of the idealized maximum-likelihood argument. No new physical or structural entities are postulated.

free parameters (4)

alpha (labeled-data classification weight) = not reported
Multiplies the supervised term log q_phi(y|x) in Eq. (4) and Algorithm 1; its value controls the balance between generative and discriminative objectives in the reported SSL results but is never given.
SVHN extra regularization weights = not reported
An unlabeled entropy confidence loss and a pseudo-loss KL term are added to the SVHN criterion in Section 4.4 with unspecified coefficients; the reported SVHN error depends on them.
inference output noise variance = 0.05 to 0.01 (GMM); 0.1 (SVHN)
Gaussian noise is added to the output of the JAE inference network; the variance schedule is hand-selected and not derived from the maximum-likelihood objective.
MIS warm-up steps = 10
Each data point's Markov chain restarts with a 10-step warm-up; the unbiasedness of the likelihood gradients depends on mixing within this budget, and no diagnostics are reported.

assumptions (5)

standard math Robbins-Monro stochastic approximation regularity conditions hold for the JSA updates
Section 3.1 invokes the SA framework, but Proposition 1 only recasts the optimum as a root-finding problem; convergence for the deep-network implementation is assumed.
domain assumption The MIS chain is at stationarity within the finite training budget
Algorithm 1 and Appendix B use a 10-step warm-up and restart per data point; posterior samples must be near p_theta(h|x) for the gradient estimates to be unbiased.
domain assumption The inference model q_phi covers the posterior support
The MIS importance ratio p_theta(h|x)/q_phi(h|x) requires q_phi to be positive wherever p_theta is positive; the inclusive KL is argued to provide this only in the infinite-capacity limit.
domain assumption The generative model family contains the true data distribution
Proposition 2 concludes p_theta*(x)=p0(x) from a zero KL; this requires the model class to be able to represent the true distribution, an unstated idealization.
standard math The empirical distribution converges to the true distribution
Used in Proposition 2; standard as n tends to infinity, but no rate or finite-sample statement is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Joint-stochastic-approximation Autoencoders with Application to Semi-supervised Learning." pith.science (2026). https://pith.science/paper/63MYTV44

@misc{pith2026250518558,
  author       = {Pith},
  title        = {Pith review of: Joint-stochastic-approximation Autoencoders with Application to Semi-supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/63MYTV44}},
  note         = {Machine review of arXiv:2505.18558}
}

read the original abstract

Our examination of existing deep generative models (DGMs), including VAEs and GANs, reveals two problems. First, their capability in handling discrete observations and latent codes is unsatisfactory, though there are interesting efforts. Second, both VAEs and GANs optimize some criteria that are indirectly related to the data likelihood. To address these problems, we formally present Joint-stochastic-approximation (JSA) autoencoders - a new family of algorithms for building deep directed generative models, with application to semi-supervised learning. The JSA learning algorithm directly maximizes the data log-likelihood and simultaneously minimizes the inclusive KL divergence the between the posteriori and the inference model. We provide theoretical results and conduct a series of experiments to show its superiority such as being robust to structure mismatch between encoder and decoder, consistent handling of both discrete and continuous variables. Particularly we empirically show that JSA autoencoders with discrete latent space achieve comparable performance to other state-of-the-art DGMs with continuous latent space in semi-supervised tasks over the widely adopted datasets - MNIST and SVHN. To the best of our knowledge, this is the first demonstration that discrete latent variable models are successfully applied in the challenging semi-supervised tasks.

Figures

Figures reproduced from arXiv: 2505.18558 by the authors.

**Figure 1.** Results for factor analysis. Upper: KL divergences between pθ(h|x) and qϕ(h|x) during training. Lower: KL divergences between the oracle p0(x) and the estimated pθ(x) during training. 4.1. Factor Analysis It is known that the encoders used in VAE training are usually not expressive enough to capture the true posterior distribution. They are often modeled as diagonal Gaussians whose means and variances are determine… view at source ↗

**Figure 3.** Column 1: Part of the training data from the context free grammar; Column 2/3/4 : data generated by JAE, GAN and VAE respectively. Both GAN and VAE use the Gumbel-softmax trick. The GAN result is copied from (Kusner & Hernandez-Lobato ´ , 2016). The temperature {0.1, 0.01, 0.001} is tested for Gumbelsoftmax with VAE. For both VAEs and JAEs, the inference network is implemented by a 3-layer fully connected neural n… view at source ↗

**Figure 2.** Comparison of a VAE with 2d Gaussian prior for latent code h (row 1), a JAE with 2d Gaussian prior (row 2) and a JAE with a mixture of 4d Bernoulli and 1d Gaussian prior (row 3). neural network, fully connected, with ReLU activations at the hidden layer and linear activations at the output layer. It can be seen from [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: Class-conditional traversal in the discrete latent space. The center images in the two pictures are the reconstructions. Surrounding images are generated with several units of the latent codes flipped randomly. The number of flipped units follows the board distance to …

**Figure 6.** Figure 6: Reconstructions (lower) for the images in the SVHN test set (upper), with 220d Bernoulli p(h). The two digits at the right of the lower picture are the true labels and predicted labels respectively [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

**Figure 7.** Figure 7: t-SNE of 60d latent codes inferred by unsupervised JAE with 60d Bernoulli p(h). Different colors represent different class variables in MNIST dataset [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

**Figure 8.** Figure 8: Architecture of the JAE model on context free grammar dataset [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

**Figure 9.** Figure 9: Architecture of the semi-JAE model on MNIST [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

**Figure 10.** Figure 10: Architecture of the semi-JAE model on SVHN [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 4 canonical work pages

[1]

Wasserstein gan

Arjovsky, M., Chintala, S., and Bottou, L. Wasserstein gan. arXiv preprint arXiv:1701.07875,

arXiv
[6]

Adversarially learned inference

Dumoulin, V ., Belghazi, I., Poole, B., Lamb, A., Arjovsky, M., Mastropietro, O., and Courville, A. Adversarially learned inference. arXiv preprint arXiv:1606.00704 ,

arXiv
[8]

J., Mnih, A., and Teh, Y

Maddison, C. J., Mnih, A., and Teh, Y . W. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016a. Maddison, C. J., Mnih, A., and Teh, Y . W. The concrete distribution: A continuous relaxation of discrete random variables. arXiv: Learning, 2016b. Makhzani, A., Shlens, J., Jaitly, N., Goodfel...

arXiv
[10]

Virtual adversarial training: a regularization method for supervised and semi-supervised learning

Miyato, T., ichi Maeda, S., Koyama, M., and Ishii, S. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. CoRR, abs/1704.03976,

arXiv
[11]

Symmetric Variational Autoencoder and Connections to Adversarial Learning

Pu, Y ., Chen, L., Dai, S., Wang, W., Li, C., and Carin, L. Symmetric variational autoencoder and connections to adversarial learning. CoRR, abs/1709.01846,

work page Pith review arXiv
[12]

O., Bousquet, O., Gelly, S., and Sch ¨olkopf, B

Tolstikhin, I. O., Bousquet, O., Gelly, S., and Sch ¨olkopf, B. Wasserstein auto-encoders. CoRR, abs/1711.01558,

arXiv
[13]

Joint Stochastic Approximation learning of Helmholtz Machines

Xu, H. and Ou, Z. Joint stochastic approximation learning of helmholtz machines. arXiv preprint arXiv:1603.06170,

work page Pith review arXiv
[1995]

Adversarial feature learning

Donahue, J., Kr ¨ahenb¨uhl, P., and Darrell, T. Adversarial feature learning. CoRR, abs/1605.09782,

arXiv

Show all 13 references

[1998]

and Bengio, Y

Bornschein, J. and Bengio, Y . Reweighted wake-sleep. arXiv preprint arXiv:1406.2751,

arXiv
[2014]

Kusner, M. J. and Hern ´andez-Lobato, J. M. Gans for se- quences of discrete elements with the gumbel-softmax distribution. CoRR, abs/1611.04051,

arXiv
[2015]

Adversar- ial variational bayes: Unifying variational autoencoders and generative adversarial networks

Mescheder, L., Nowozin, S., and Geiger, A. Adversar- ial variational bayes: Unifying variational autoencoders and generative adversarial networks. arXiv preprint arXiv:1701.04722,

arXiv
[2016]

Importance weighted autoencoders

Burda, Y ., Grosse, R., and Salakhutdinov, R. Importance weighted autoencoders. arXiv preprint arXiv:1509.00519,

arXiv
[2017]

Bengio, Y ., L´eonard, N., and Courville, A. C. Estimating or propagating gradients through stochastic neurons for conditional computation. CoRR, abs/1308.3432,

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Wasserstein gan

Arjovsky, M., Chintala, S., and Bottou, L. Wasserstein gan. arXiv preprint arXiv:1701.07875,

arXiv

[2] [6]

Adversarially learned inference

Dumoulin, V ., Belghazi, I., Poole, B., Lamb, A., Arjovsky, M., Mastropietro, O., and Courville, A. Adversarially learned inference. arXiv preprint arXiv:1606.00704 ,

arXiv

[3] [8]

J., Mnih, A., and Teh, Y

Maddison, C. J., Mnih, A., and Teh, Y . W. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016a. Maddison, C. J., Mnih, A., and Teh, Y . W. The concrete distribution: A continuous relaxation of discrete random variables. arXiv: Learning, 2016b. Makhzani, A., Shlens, J., Jaitly, N., Goodfel...

arXiv

[4] [10]

Virtual adversarial training: a regularization method for supervised and semi-supervised learning

Miyato, T., ichi Maeda, S., Koyama, M., and Ishii, S. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. CoRR, abs/1704.03976,

arXiv

[5] [11]

Symmetric Variational Autoencoder and Connections to Adversarial Learning

Pu, Y ., Chen, L., Dai, S., Wang, W., Li, C., and Carin, L. Symmetric variational autoencoder and connections to adversarial learning. CoRR, abs/1709.01846,

work page Pith review arXiv

[6] [12]

O., Bousquet, O., Gelly, S., and Sch ¨olkopf, B

Tolstikhin, I. O., Bousquet, O., Gelly, S., and Sch ¨olkopf, B. Wasserstein auto-encoders. CoRR, abs/1711.01558,

arXiv

[7] [13]

Joint Stochastic Approximation learning of Helmholtz Machines

Xu, H. and Ou, Z. Joint stochastic approximation learning of helmholtz machines. arXiv preprint arXiv:1603.06170,

work page Pith review arXiv

[8] [1995]

Adversarial feature learning

Donahue, J., Kr ¨ahenb¨uhl, P., and Darrell, T. Adversarial feature learning. CoRR, abs/1605.09782,

arXiv