Half-AVAE: Adversarial-Enhanced Factorized and Structured Encoder-Free VAE for Underdetermined Independent Component Analysis

Yan-Jie Sun; Yuan-Hao Wei

REVIEW 3 major objections 5 minor 36 references

Half-AVAE: Adversarial-Enhanced Factorized and Structured Encoder-Free VAE for Underdetermined Independent Component Analysis

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 Half-AVAE claims to solve underdetermined independent component analysis by dropping the encoder entirely and instead enforcing latent independence with an adversarial discriminator and prior-separation terms, cutting average RMSE to…

desk verdict Plausible architecture, unsupported headline result: the underdetermined-ICA gain is a single matched-prior synthetic run with no seeds, baselines, or sensitivity analysis. read the letter →

arxiv 2506.07011 v1 pith:TEWMC2JF submitted 2025-06-08 stat.ML cs.LGeess.SP

Yuan-Hao Wei , Yan-Jie Sun This is my paper

classification stat.MLcs.LGeess.SP MSC 62H2568T07

keywords underdeterminedICAencoder-freeVAEadversarialindependenceGaussianprocesspriorblindsourceseparationvariationalinferencedisentanglementexternalenhancement

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 claims that underdetermined independent component analysis — recovering more latent source signals than there are observed mixtures — can be solved by a variational autoencoder that simply drops the encoder. Its model, Half-AVAE, treats the approximate posterior as a set of directly optimized distribution parameters rather than the output of a neural encoder, and adds two mechanisms to force the latent dimensions apart: an adversarial discriminator that aligns the joint distribution of latent sequences with the product of their marginals, and External Enhancement terms that push the Gaussian-process prior length scales of different dimensions to diverge. On a synthetic three-source, two-observation task, the method reports an average RMSE of 0.2272, against 0.6474 for the adversarial encoder-based GP-AVAE and 0.5756 for the encoder-free Half-GP-VAE without the enhancement terms. If correct, this shows the encoder is not essential to VAEs, and that variational inference with structured priors plus explicit independence pressure is enough to make underdetermined ICA tractable.

What carries the argument

The load-bearing mechanism is the encoder-free variational objective combined with distribution-level independence enforcement. Instead of an encoder mapping X to q(Z|X), the model directly optimizes the parameters Ω=(μz,$σ^{2}$_z) of a distribution IΩ(Z) to minimize −E_{IΩ(Z)}[ln PΨ(X|Z)] + KL(IΩ(Z)‖PΓ(Z)), so no inverse mapping $f^{{-1}}$ needs to exist. Independence is enforced two ways: an adversarial network whose discriminator DΦ distinguishes joint samples {Z^a_τ,…,Z^n_τ} from shuffled marginal samples, pulling the joint distribution R(Z) toward the product ∏R(Z^i); and External Enhancement terms that separate the GP-SE prior length scales Γ_i and penalize variances, making each latent dimension a distinct structured temporal process.

What would settle it

Run the identical underdetermined three-source, two-mixture experiment with sources that are not squared-exponential Gaussian processes — for example independent Laplace or AR(1) signals with known length scales — and check whether Half-GP-AVAE still achieves RMSE near 0.23 rather than degrading toward the Half-GP-VAE level; alternatively, ablate the EE terms (β1=β2=β3=0) and verify whether the adversarial module alone sustains the improvement.

Watch

Extended reading notes

Core claim

The central discovery is that an encoder-free VAE with factorized Gaussian-process priors, adversarial independence training, and External Enhancement terms recovers independent components in underdetermined linear ICA. The paper constructs Half-GP-AVAE by taking the Half-VAE posterior IΩ(Z), whose mean sequence and variance for each latent dimension are optimized directly through the ELBO, and coupling it with a discriminator that treats jointly sampled latent tuples as real and shuffled (marginally matched) tuples as fake, thereby driving the joint distribution toward the product of the marginal distributions. External Enhancement terms β1 Σ 1/(Γi−Γj)^2 + β2 Σ Γi + β3 Σ $σ^{2}$_{Zi} are added to force distinct prior length scales and bounded variances. In the underdetermined experiment with three GP-SE-structured sources and two mixtures, the method achieves average RMSE 0.2272 versus 0.6474 (GP-AVAE) and 0.5756 (Half-GP-VAE), and the paper reads this as validating the theoretical correctness of the Half-VAE framework for underdetermined ICA.

Load-bearing premise

The experiments generate sources as independent Gaussian processes with squared-exponential kernels, the same family the model uses as its prior, so the reported gains may depend on a matched prior-data setup rather than a general ability to solve underdetermined ICA.

Editorial extensions

If this is right

Underdetermined linear mixtures of structured sources can be separated without any encoder network, as long as the posterior is optimized directly and independence is pressured explicitly.
Adversarial marginal-joint alignment plus prior-length-scale separation together give a large error reduction (Table 2: 0.2272 vs. 0.5756/0.6474), implying each mechanism alone leaves a gap.
The results suggest that VAE-based ICA pipelines should be built around the variational bound and prior design rather than the encoder-decoder symmetry.
The method's success on time-structured sources points toward applications in blind source separation, disentanglement, and causal representation learning where n > m.

Reading between the lines

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

Because the synthetic sources are drawn from the same GP-SE family used as the prior, the reported advantage may narrow on sources with different temporal structure; a natural test is i.i.d. or heavy-tailed sources.
The EE hyperparameters β1, β2, β3 are hand-set without sensitivity analysis, so transfer to new tasks may require retuning; a principled schedule or automatic selection would strengthen the approach.
The shuffling-based adversarial independence estimator could be replaced by other dependence measures (such as HSIC or distance covariance) with the same encoder-free backbone, yielding a family of underdetermined ICA solvers.
If the method generalizes, it offers a practical route for neuroimaging or audio source separation where sensors are scarce relative to sources.

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

Let me give you the short version: the underdetermined-ICA result in this paper is not yet believable, but that is not because the idea is silly. The new bit is combining the encoder-free Half-VAE with adversarial independence training and three hand-tuned external-enhancement terms, and the determined-case sanity check is a reasonable first step. The problem is that the experimental support for the headline number (RMSE 0.2272 vs 0.58–0.65) is a single synthetic run with no seeds, no error bars, no standard ICA baselines, and no code or data. I agree with the stress-test note: the concern holds up on reading. Section 3.1 never says how the synthetic sources were generated. It says only that they have 'distinct temporal structures', then defines a GP-SE prior 'to capture these structures'. If the sources were sampled from the same GP-SE family, the experiment is a matched-prior test: the model is being shown data that exactly matches its inductive bias. That makes the comparison with GP-AVAE and Half-GP-VAE less damning—all three share the prior, but Half-AVAE is the one that most aggressively fits it, so it can win at an in-model fitting game. And the EE weights were tuned on this same task, so without sensitivity analysis the margin in Table 2 could be selection on the test set. Section 4's claim that this validates the 'theoretical correctness' of the Half-VAE framework goes well beyond what a single favorable run can support. What the paper does well: the derivation of the ELBO and the adversarial independence mechanism is standard but clear, and the architecture is described enough to reproduce if you had the code. The direction is genuinely interesting—encoder-free variational inference for underdetermined ICA with adversarial independence is worth testing properly. It just has not been tested here. To be credible, the paper needs sources outside the GP-SE family, multiple random seeds with error bars, standard baselines like FastICA, and a sensitivity analysis for the EE weights. The citation pattern leans on the authors' prior work, but that is not the issue; the missing evidence is. My recommendation: not publishable as is, but do not desk-reject it. Send it to a referee with a clear request to focus on the experimental design. If the result survives a real out-of-prior test, it would be a useful incremental contribution. If not, the authors need to find another way to demonstrate the value of encoder-free inference.

Referee Report

3 major / 5 minor

Summary. This paper proposes Half-AVAE, an encoder-free variational autoencoder that combines Gaussian Process priors with an adversarial discriminator and External Enhancement (EE) regularizers to perform underdetermined independent component analysis. The method is derived from the standard VAE ELBO and extends the authors' prior Half-VAE and GP-AVAE frameworks. Experiments on synthetic signals compare Half-GP-AVAE against GP-AVAE and Half-GP-VAE under determined and underdetermined conditions, reporting lower RMSE for Half-GP-AVAE in the underdetermined case (Table 2). The paper concludes that these results validate the theoretical correctness of the Half-VAE framework for underdetermined ICA.

Significance. If the reported empirical result is reproducible, the proposed combination is a plausible way to handle underdetermined linear ICA for structured time-series sources, and the paper is clearly written at the conceptual level. The derivation of the ELBO in Section 2 is standard, but the concrete combination of encoder-free variational inference, adversarial independence, and EE regularizers is coherent and the synthetic comparison is a useful sanity check. However, the absence of code, the underspecified data-generation process, and the single-run evaluation substantially limit the significance as presented.

major comments (3)

[Section 3.1] The generative process of the synthetic sources is not specified; the paper only states that three sequences with 'distinct temporal or spatial structures' were generated, and then defines the GP-SE prior in Eq. (15) 'to capture these structures.' Without stating whether the source sequences were sampled from independent GP-SE processes or generated in another way, the reader cannot exclude a matched-prior setting in which the model's prior is the true data-generating process. This is load-bearing because the central empirical claim (Section 3.3.2, Table 2) is the superior RMSE of Half-GP-AVAE; the authors should provide the exact generative model or code.
[Section 3.3.2 and Table 2] The reported RMSE values come from a single run with no error bars, no number of random seeds, no train/validation/test split, and no sensitivity analysis for the EE weights β1, β2, β3 in Eq. (16) or the adversarial weight Λ in Eq. (14a). The text explicitly says these weights were 'tuned to balance' the loss terms on this same synthetic task, so the large margin (0.2272 vs. 0.5756/0.6474) may reflect selection on the test set rather than a genuine algorithmic improvement. Repeated runs with multiple seeds and a hold-out protocol are needed.
[Section 4] The statement that the results 'validate the theoretical correctness of the Half-VAE framework' is an overclaim. The experiments are a single synthetic demonstration; they do not provide identifiability guarantees or a theoretical argument for underdetermined ICA recovery. The conclusion should be rephrased as an empirical feasibility result, or the paper should add a formal identifiability analysis.

minor comments (5)

[Sections 2.3 and 2.4] Equation (6) is numbered twice: once for the factorization of the prior and once for the independence condition used in the adversarial section. Renumber to avoid confusion.
[Section 2.4] The notation R(Zi) is used ambiguously: the text says 'the sequence Zi 1:T corresponds to qΘ(Zi|X)' but the adversarial module operates on the mean sequence μZi 1:T. Please clarify the distinction between the distribution over latent sequences and the distribution over mean sequences.
[Tables 1 and 2] The table captions do not specify that the reported values are RMSE after z-score normalization; this information appears only in the running text. Add it to the captions for readability.
[Section 3.3.2] The sentence 'The EE terms yield minimal improvement in GP-AVAE's ICA performance' could be quantified by reporting the actual RMSE differences from Table 2 rather than leaving the reader to compute them.
[References] Reference 'Wei, Y. et al. (2024a)' is cited as an unpublished thesis or report; please provide a published, accessible version or more bibliographic detail so readers can verify the prior work.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; the encoder-free derivation is algebraically self-contained, and the matched-prior concern is an experimental-design limitation rather than a circular step.

full rationale

The paper's derivation chain is self-contained: Equations (4)-(5) give the standard VAE ELBO, and Equation (13) replaces the encoder posterior q(Z|X) with a directly optimized distribution IOmega(Z). This replacement is algebraically valid because the ELBO lower bound holds for any variational distribution, so the encoder-free step does not reduce by construction to the prior work it cites. The adversarial objective (Eq. 9) and the EE regularizer (Eq. 16) are added to that ELBO, and the reported RMSE values in Tables 1-2 are empirical outcomes of optimizing the resulting losses, not quantities fixed by the fitted parameters themselves. The self-citations to Wei et al. (2024a,b) motivate the GP priors and Half-VAE architecture, but the central encoder-free claim is also supported by the paper's own equations and by the experiments, so no uniqueness theorem or unverified self-citation chain is load-bearing. The main caveat is in Section 3.1, where the paper says the GP-SE prior is designed 'to capture these structures' of the synthetic sources without specifying the generative distribution of those sources; if the sources were sampled from the same GP-SE family, the experiment would be a matched-prior sanity check rather than an external validation. However, the paper never states that the sources were drawn from Eq. (15), the baselines share the same prior, and the reported advantage of Half-GP-AVAE is not forced by construction. Missing seeds, error bars, and sensitivity analysis affect robustness, not circularity. Therefore no specific circular step is established.

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

The core derivation is standard variational inference. The load-bearing extra assumptions are domain-specific: that encoder-free direct optimization of Z is a valid posterior approximation for underdetermined ICA, and that the sources live in the same GP-SE family the model uses as a prior. The method adds no new physical or mathematical entities, but it does add hand-tuned loss terms that are fitted to the single experiment.

free parameters (4)

EE weights β1, β2, β3 = not reported
Hyperparameters in Eq. (16) that push prior length scales apart and penalize unbounded growth. They are hand-tuned on the same synthetic task and no sensitivity analysis is given.
Λ adversarial weight = not reported
Hyperparameter in Eqs. (14a) and (19a) that balances the adversarial loss against the VAE objective; the value is not reported.
GP length scales Γi = trainable, not reported
Per-dimension length-scale parameters of the GP priors in Eq. (15). They are optimized jointly with the model, blurring the boundary between prior specification and fitted parameters.
Latent variance parameters σ²_Zi = trainable, not reported
Per-sequence variances in the inferred distribution IΩ(Z), regularized by the third EE term. Their fitted values are not reported.

assumptions (4)

standard math The ELBO objective is a valid surrogate for matching the variational distribution to the posterior (Sec. 2.1).
Standard variational inference result used to derive the training objective.
domain assumption The true sources are independent Gaussian processes with squared-exponential kernels (Sec. 3.1).
The synthetic experiment generates sources with smooth temporal structure, and the model uses the same GP-SE prior, making the data matched to the prior family.
domain assumption An encoder-free variational distribution IΩ(Z) can approximate the true posterior and thereby solve underdetermined ICA (Sec. 2.5).
This assertion is carried by a citation to the authors' own Half-VAE paper; no identifiability or consistency proof is given.
standard math Shuffling the index sequence preserves each marginal distribution R(Zi), so shuffling yields samples from the product of marginals (Sec. 2.4).
Probabilistic identity used to construct adversarial training pairs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Half-AVAE: Adversarial-Enhanced Factorized and Structured Encoder-Free VAE for Underdetermined Independent Component Analysis." pith.science (2026). https://pith.science/paper/TEWMC2JF

@misc{pith2026250607011,
  author       = {Pith},
  title        = {Pith review of: Half-AVAE: Adversarial-Enhanced Factorized and Structured Encoder-Free VAE for Underdetermined Independent Component Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TEWMC2JF}},
  note         = {Machine review of arXiv:2506.07011}
}

read the original abstract

This study advances the Variational Autoencoder (VAE) framework by addressing challenges in Independent Component Analysis (ICA) under both determined and underdetermined conditions, focusing on enhancing the independence and interpretability of latent variables. Traditional VAEs map observed data to latent variables and back via an encoder-decoder architecture, but struggle with underdetermined ICA where the number of latent variables exceeds observed signals. The proposed Half Adversarial VAE (Half-AVAE) builds on the encoder-free Half-VAE framework, eliminating explicit inverse mapping to tackle underdetermined scenarios. By integrating adversarial networks and External Enhancement (EE) terms, Half-AVAE promotes mutual independence among latent dimensions, achieving factorized and interpretable representations. Experiments with synthetic signals demonstrate that Half-AVAE outperforms baseline models, including GP-AVAE and Half-VAE, in recovering independent components under underdetermined conditions, as evidenced by lower root mean square errors. The study highlights the flexibility of VAEs in variational inference, showing that encoder omission, combined with adversarial training and structured priors, enables effective solutions for complex ICA tasks, advancing applications in disentanglement, causal inference, and generative modeling.

Figures

Figures reproduced from arXiv: 2506.07011 by the authors.

**Figure 2.** Adversarial Network Module Integrated with VAE Architecture for Enhancing Independence [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Half-AVAE 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Synthetic Sources 9 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: illustrates the observations X, obtained via the mixing mapping f, i.e., X = f(Z) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: ICA Results of Determined Scenario GP-AVAE Half-GP-VAE Half-GP-AVAE Independent component 1 0.8494 0.8468 0.6981 Independent component 2 0.4021 0.1736 0.3317 Independent component 3 0.6941 0.7256 0.5185 Average 0.6485 0.582 0.5161 [PITH_FULL_IMAGE:figures/full_fig_p01…

**Figure 7.** Figure 7: Observations of Underdetermined Scenario [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: ICA Results of Underdetermined Scenario pendence among the dimensions Z = {Z a , Z b , . . . , Z n}. The EE terms are defined as: EE = β1 X 1≤i<j≤n 1 (Γi − Γj ) 2 + β2 Xn i=1 Γ i + β3 Xn i=1 σ 2 Zi . (16) The first term encourages the prior distribution parameters Γi t…

**Figure 9.** Figure 9: ICA Results of Underdetermined Scenario 4 Discussion and Conclusion Theoretically, Variational Autoencoders (VAEs) with explicit encoders cannot perform the inverse mapping required for Independent Component Analysis (ICA) under underdetermined conditions, where the n…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 30 canonical work pages

[1]

Ahuja, K., Mahajan, D., Wang, Y., and Bengio, Y. (2023). Interventional causal representation learning. In International conference on machine learning , pages 372--407. PMLR

work page 2023
[2]

Arjovsky, M., Chintala, S., and Bottou, L. (2017). Wasserstein generative adversarial networks. In International conference on machine learning , pages 214--223. PMLR

work page 2017
[3]

and Kamp, Y

Bourlard, H. and Kamp, Y. (1988). Auto-association by multilayer perceptrons and singular value decomposition. Biological cybernetics , 59(4):291--294

work page 1988
[4]

Learning Independent Features with Adversarial Nets for Non-linear ICA

Brakel, P. and Bengio, Y. (2017). Learning independent features with adversarial nets for non-linear ica. arXiv preprint arXiv:1710.05050

work page Pith review arXiv 2017
[5]

P., Higgins, I., Pal, A., Matthey, L., Watters, N., Desjardins, G., and Lerchner, A

Burgess, C. P., Higgins, I., Pal, A., Matthey, L., Watters, N., Desjardins, G., and Lerchner, A. (2018). Understanding disentangling in -vae. arXiv preprint arXiv:1804.03599

arXiv 2018
[6]

Cardoso, J.-F. (1998). Blind signal separation: statistical principles. Proceedings of the IEEE , 86(10):2009--2025

work page 1998
[7]

T., Li, X., Grosse, R

Chen, R. T., Li, X., Grosse, R. B., and Duvenaud, D. K. (2018). Isolating sources of disentanglement in variational autoencoders. Advances in neural information processing systems , 31

work page 2018
[8]

Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., and Abbeel, P. (2016). Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems , 29

work page 2016

Show all 36 references

[9]

Comon, P. (1994). Independent component analysis, a new concept? Signal processing , 36(3):287--314

1994
[10]

M., Mircea, M., Mueller, N

Eraslan, G., Simon, L. M., Mircea, M., Mueller, N. S., and Theis, F. J. (2019). Single-cell rna-seq denoising using a deep count autoencoder. Nature communications , 10(1):390

2019
[11]

Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., and Lempitsky, V. (2016). Domain-adversarial training of neural networks. Journal of machine learning research , 17(59):1--35

2016
[12]

J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems , 27

2014
[13]

H., Vording, M

Gr nbech, C. H., Vording, M. F., Timshel, P. N., S nderby, C. K., Pers, T. H., and Winther, O. (2020). scvae: variational auto-encoders for single-cell gene expression data. Bioinformatics , 36(16):4415--4422

2020
[14]

P., Glorot, X., Botvinick, M

Higgins, I., Matthey, L., Pal, A., Burgess, C. P., Glorot, X., Botvinick, M. M., Mohamed, S., and Lerchner, A. (2017). beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR (Poster) , 3

2017
[15]

and Glass, J

Hsu, W.-N. and Glass, J. (2018). Scalable factorized hierarchical variational autoencoder training. arXiv preprint arXiv:1804.03201

2018 arXiv
[16]

Hyvarinen, A., Karhunen, J., and Oja, E. (2001). Independent component analysis and blind source separation

2001
[17]

Hyvarinen, A., Sasaki, H., and Turner, R. (2019). Nonlinear ica using auxiliary variables and generalized contrastive learning. In The 22nd International Conference on Artificial Intelligence and Statistics , pages 859--868. PMLR

2019
[18]

Khemakhem, I., Kingma, D., Monti, R., and Hyvarinen, A. (2020). Variational autoencoders and nonlinear ica: A unifying framework. In International conference on artificial intelligence and statistics , pages 2207--2217. PMLR

2020
[19]

Kingma, D. P. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

2013 arXiv
[20]

Kingma, D. P. and Welling, M. (2019). An introduction to variational autoencoders. arxiv e-prints, page. arXiv preprint arXiv:1906.02691

2019 arXiv
[21]

Komanduri, A., Wu, Y., Huang, W., Chen, F., and Wu, X. (2022). Scm-vae: Learning identifiable causal representations via structural knowledge. In 2022 IEEE International Conference on Big Data (Big Data) , pages 1014--1023. IEEE

2022
[22]

E., Le Priol, R., Lacoste, A., and Lacoste-Julien, S

Lachapelle, S., Rodriguez, P., Sharma, Y., Everett, K. E., Le Priol, R., Lacoste, A., and Lacoste-Julien, S. (2022). Disentanglement via mechanism sparsity regularization: A new principle for nonlinear ica. In Conference on Causal Learning and Reasoning , pages 428--484. PMLR

2022
[23]

Liang, T. (2021). How well generative adversarial networks learn distributions. Journal of Machine Learning Research , 22(228):1--41

2021
[24]

M., Cohen, T., and Gavves, S

Lippe, P., Magliacane, S., L \"o we, S., Asano, Y. M., Cohen, T., and Gavves, S. (2022). Citris: Causal identifiability from temporal intervened sequences. In International Conference on Machine Learning , pages 13557--13603. PMLR

2022
[25]

B., Jordan, M

Lopez, R., Regier, J., Cole, M. B., Jordan, M. I., and Yosef, N. (2018). Deep generative modeling for single-cell transcriptomics. Nature methods , 15(12):1053--1058

2018
[26]

Pearl, J. (2019). The seven tools of causal inference, with reflections on machine learning. Communications of the ACM , 62(3):54--60

2019
[27]

J., Mohamed, S., and Wierstra, D

Rezende, D. J., Mohamed, S., and Wierstra, D. (2014). Stochastic backpropagation and approximate inference in deep generative models. In International conference on machine learning , pages 1278--1286. PMLR

2014
[28]

E., Hinton, G

Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning representations by back-propagating errors. nature , 323(6088):533--536

1986
[29]

Sch \"o lkopf, B. (2022). Causality for machine learning. In Probabilistic and causal inference: The works of Judea Pearl , pages 765--804

2022
[30]

Tzeng, E., Hoffman, J., Saenko, K., and Darrell, T. (2017). Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 7167--7176

2017
[31]

and Gu, J

Wang, D. and Gu, J. (2018). Vasc: dimension reduction and visualization of single-cell rna-seq data by deep variational autoencoder. Genomics, proteomics & bioinformatics , 16(5):320--331

2018
[32]

Way, G. P. and Greene, C. S. (2018). Extracting a biologically relevant latent space from cancer transcriptomes with variational autoencoders. In PACIFIC SYMPOSIUM on BIOCOMPUTING 2018: Proceedings of the Pacific Symposium , pages 80--91. World Scientific

2018
[33]

Wei, Y. et al. (2024a). Innovative blind source separation techniques combining gaussian process algorithms and variational autoencoders with applications in structural health monitoring

2024
[34]

Wei, Y.-H. (2025). Vaes and gans: Implicitly approximating complex distributions with simple base distributions and deep neural networks--principles, necessity, and limitations. arXiv preprint arXiv:2503.01898

2025 arXiv
[35]

Wei, Y.-H., Sun, Y.-J., and Zhang, C. (2024b). Half-vae: An encoder-free vae to bypass explicit inverse mapping. arXiv preprint arXiv:2409.04140

2024 arXiv
[36]

Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., and Wang, J. (2021). Causalvae: Disentangled representation learning via neural structural causal models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9593--9602

2021

Pith tools

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

[1] [1]

Ahuja, K., Mahajan, D., Wang, Y., and Bengio, Y. (2023). Interventional causal representation learning. In International conference on machine learning , pages 372--407. PMLR

work page 2023

[2] [2]

Arjovsky, M., Chintala, S., and Bottou, L. (2017). Wasserstein generative adversarial networks. In International conference on machine learning , pages 214--223. PMLR

work page 2017

[3] [3]

and Kamp, Y

Bourlard, H. and Kamp, Y. (1988). Auto-association by multilayer perceptrons and singular value decomposition. Biological cybernetics , 59(4):291--294

work page 1988

[4] [4]

Learning Independent Features with Adversarial Nets for Non-linear ICA

Brakel, P. and Bengio, Y. (2017). Learning independent features with adversarial nets for non-linear ica. arXiv preprint arXiv:1710.05050

work page Pith review arXiv 2017

[5] [5]

P., Higgins, I., Pal, A., Matthey, L., Watters, N., Desjardins, G., and Lerchner, A

Burgess, C. P., Higgins, I., Pal, A., Matthey, L., Watters, N., Desjardins, G., and Lerchner, A. (2018). Understanding disentangling in -vae. arXiv preprint arXiv:1804.03599

arXiv 2018

[6] [6]

Cardoso, J.-F. (1998). Blind signal separation: statistical principles. Proceedings of the IEEE , 86(10):2009--2025

work page 1998

[7] [7]

T., Li, X., Grosse, R

Chen, R. T., Li, X., Grosse, R. B., and Duvenaud, D. K. (2018). Isolating sources of disentanglement in variational autoencoders. Advances in neural information processing systems , 31

work page 2018

[8] [8]

Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., and Abbeel, P. (2016). Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems , 29

work page 2016