Pith. sign in

REVIEW 3 major objections 6 minor 34 references

Bidirectional Variational Autoencoders

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single network can do both encoding and decoding in a variational autoencoder, cutting parameters nearly in half without losing performance.

desk verdict A tied-weight VAE with a half-parameter architecture and mixed empirical evidence; the theory is standard ELBO, the experiments need seeds and error bars. read the letter →

arxiv 2505.16074 v2 pith:WZM3ROB3 submitted 2025-05-21 cs.LG cs.AIstat.ML

Bart Kosko , Olaoluwa Adigun This is my paper · ORCID
classification cs.LGcs.AIstat.ML
keywords bidirectionalvariationalautoencoderevidencelowerboundbackpropagationdirectionallikelihoodsimagegenerationweighttyingrepresentationlearning
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 proposes the bidirectional variational autoencoder (BVAE), which replaces the usual two-network encoder-decoder setup of a variational autoencoder with one network that encodes on its forward pass and decodes on its backward pass through the same synaptic weights. Training maximizes the BELBO, a bidirectional version of the evidence lower bound, using a bidirectional backpropagation algorithm that jointly optimizes the forward and backward directional likelihoods. The paper reports that on MNIST, Fashion-MNIST, CIFAR-10, and CelebA-64, BVAEs use almost 50% fewer parameters than comparable unidirectional VAEs and still perform slightly better on reconstruction, classification, interpolation, and generation. If the central claim holds, VAE training and deployment would cost roughly half as much in memory and parameter storage for equal or better quality.

What carries the argument

The central mechanism is the BELBO (bidirectional evidence lower bound), built on the directional likelihoods of bidirectional backpropagation: the forward likelihood $q_f(z|x,\theta)$ encodes the latent variable and the backward likelihood $p_b(x|z,\theta)$ decodes it through the same synaptic web. The bound is obtained by inserting these two densities into the identity for the data log-likelihood and discarding the nonnegative KL divergence between $q_f$ and the true posterior, which turns the intractable marginal likelihood into a tractable one-parameter objective. The BVAE training loop estimates this bound with the reparameterization trick for the latent sample and updates the shared weight vector with ordinary backpropagation through the BELBO estimate.

What would settle it

Compute an unbiased Monte Carlo estimate of the true held-out log-likelihood for a trained BVAE on a standard dataset and check whether the BELBO value reported by the same model exceeds that estimate. If the BELBO is larger than the true log-likelihood on any sample, the claimed bound fails. A second check is to train a tied-weight autoencoder with the same architecture but no variational objective; if it matches the BVAE on all tasks, the variational machinery adds nothing beyond weight tying.

Watch

Extended reading notes

Core claim

The central claim is that the encoder and decoder of a variational autoencoder need not be separate networks: a single parameterized network can serve both roles if it is trained with bidirectional backpropagation. In a BVAE, the forward likelihood $q_f(z|x,\theta)$ approximates the intractable posterior $q(z|x,\theta)$ and the backward likelihood $p_b(x|z,\theta)$ approximates the decoder likelihood $p(x|z,\theta)$, with both densities sharing the same weight vector $\theta$. The paper derives the BELBO by substituting these directional likelihoods into the standard log-likelihood identity and dropping a nonnegative KL divergence term, giving a one-parameter objective that the network maximizes. The simulations compare vanilla VAEs, $\beta$-VAEs, $\beta$-TCVAEs, and IWAEs with their bidirectional counterparts and report that the bidirectional versions halve the parameter count while slightly outperforming the unidirectional ones on the tested image tasks.

Load-bearing premise

The whole variational argument depends on the backward-pass density $p_b(x|z,\theta)$ actually equaling the true decoder likelihood $p(x|z,\theta)$ for the data distribution; if it does not, the BELBO is not a lower bound on the log-likelihood and the architecture is just a tied-weight autoencoder whose performance rests on the simulations.

Editorial extensions

If this is right

  • If the central claim is correct, a variational autoencoder can be trained and run with a single network, cutting parameter storage and memory roughly in half while maintaining or slightly improving image reconstruction, classification, generation, and interpolation performance.
  • The same shared-weight substitution applies to VAE variants such as $\beta$-VAE, $\beta$-TCVAE, and IWAE, so the parameter savings should carry over to any objective built on the ELBO.
  • The backward pass uses transposed convolution with the same masks as the forward convolution, so the architecture enforces a form of weight tying that acts as a regularizer and may explain the small generalization gains reported in the simulations.
  • Because the savings scale with the number of trainable weights, deeper and wider autoencoders would benefit more from the bidirectional architecture than the relatively shallow networks tested here.

Reading between the lines

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

  • A direct test of the variational claim would be to compare the BELBO value reported during training against an unbiased Monte Carlo estimate of the true held-out log-likelihood; if the BELBO ever exceeds that estimate, the bound is not a true lower bound and the method's value rests on the empirical results alone.
  • The backward-pass density $p_b(x|z,\theta)$ is the transpose-weight mapping, not a freely parameterized decoder; datasets with strongly nonlinear conditional structure might show a larger performance gap because the decoder's capacity is constrained by the encoder's weights.
  • The same bidirectional weight-sharing idea could be applied to other latent-variable generative models, such as hierarchical VAEs or normalizing flows, whenever a backward pass can approximate the inverse of the forward map.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes a bidirectional variational autoencoder (BVAE) in which a single network is used both as the encoder, through a forward pass with likelihood q_f(z|x,θ), and as the decoder, through a backward pass with likelihood p_b(x|z,θ). The authors derive a lower bound they call BELBO, which parallels the standard ELBO, and report experiments on MNIST, Fashion-MNIST, CIFAR-10, and CelebA-64 comparing vanilla VAEs, β-VAEs, β-TCVAEs, and IWAEs with their bidirectional counterparts across reconstruction, classification, interpolation, and generation. The central claims are that the bidirectional architecture approximately halves the parameter count and that it still slightly outperforms unidirectional VAEs.

Significance. If the empirical claims were established, the paper would offer a simple and parameter-efficient VAE variant with tied encoder/decoder weights, which is of practical interest. The paper deserves credit for presenting a concrete architecture and for a mostly correct review of the standard ELBO derivation in Section II. However, the theoretical contribution is currently overstated: once the backward-pass density p_b is properly declared as the generative decoder, BELBO is the standard ELBO for the tied-weight model, not a new lower bound on the true data likelihood. The headline empirical claim is also not supported by the tables as reported, because each configuration appears to be a single run with no error bars or significance tests and the results are inconsistent across metrics and model families. The paper does not provide code, random seeds, or detailed training configurations, which further limits reproducibility.

major comments (3)
  1. [III, Eqs. (26)–(38) and Algorithm 1, Step 8] The BELBO derivation is not yet a proof that the training objective bounds the data log-likelihood. Eqs. (26)–(38) are written with p(x|z,θ) as the decoder, but Algorithm 1 and the surrounding text replace this density with the backward-pass likelihood p_b(x|z,θ). The manuscript never defines the generative model p_model(x)=∫ p_b(x|z,θ)p(z|θ)dz, nor does it state that the inequality in Eq. (37) is intended for that model. If p_b is taken as the decoder likelihood, Eq. (38) is simply the standard ELBO for a tied-weight model, and the derivation is complete but not novel; if p_b is not the true conditional likelihood, then no correction term is provided and the bound on ln p(x|θ) is unproven. Please state the model explicitly and adjust the claim accordingly.
  2. [Tables I–IV and Abstract] The abstract's claim that BVAEs 'slightly outperformed' unidirectional VAEs is not established by the reported experiments. No random seeds, repeated runs, error bars, or paired significance tests are given, so the observed differences may be initialization noise. The results are also internally inconsistent across metrics: in Table I the vanilla BVAE has better NLL (86.69 vs 86.72) but worse FID (3.370 vs 3.340), and the bidirectional IWAE has better NLL (86.00 vs 86.02) but worse FID (3.330 vs 2.958); in Table II the bidirectional IWAE has worse FID (2.881 vs 2.489) and worse accuracy (88.05% vs 88.14%); in Table IV the bidirectional β-VAE has substantially worse NLL (6261.9 vs 6243.0). Either provide multiple seeds with standard deviations and significance tests, or scale back the headline claim to one of parameter-count reduction with comparable performance on selected metrics.
  3. [I, Eqs. (2)–(3), and III, Algorithm 1] The relationship between the bidirectional backpropagation objective and the BELBO used for training is not specified. Eq. (2) defines B-BP as maximizing q_f(z|x,θ)p_b(x|z,θ), equivalently maximizing ln q_f(z|x,θ)+ln p_b(x|z,θ). Algorithm 1 maximizes BCE+KLD, i.e., E_{q_f}[ln p_b(x|z,θ)] - KL(q_f||p(z)), which is a different functional: the former contains no expectation over the variational distribution and no prior KL term. The text cites [2] for B-BP but does not show that B-BP implements BELBO maximization. Please clarify how the update in Eq. (39) is derived from the B-BP algorithm, or present Algorithm 1 as the definition of BVAE training and describe its relation to Eq. (2).
minor comments (6)
  1. [Algorithm 1, Step 8] The estimated objective is written as \tilde{L}_{ELBO}(x,θ), but in the BVAE setting it should be \tilde{L}_{BELBO}(x,θ) to match the notation of Section III.
  2. [II, end of Section II] The sentence 'the forward pass p_f(x|z,θ) to approximate the encoding model' should read 'decoding model'; p_f(x|z,θ) is a decoder likelihood, not an encoder.
  3. [II, Eqs. (16)–(21)] The sentence 'because the negative of the natural logarithm is convex' is imprecise; the argument applies Jensen's inequality to the concave logarithm and then uses the nonnegativity of the negative of that inequality. Please rephrase.
  4. [Tables I–IV] Several numerical entries are difficult to read because of spacing artifacts, e.g., '18 4 .340' in Table I and '2 .483' in Table III. Please reformat the tables so that every number is unambiguous.
  5. [References] References [6] and [7] are the same arXiv paper and ICLR paper by Burda, Grosse, and Salakhutdinov; they should be consolidated into a single citation.
  6. [IV-E, Eq. (41)] The active-units metric should be defined as the variance over the data of the posterior mean, Var_x(E_{z|x}[z_d]), not 'Cov_x' as written, and the threshold should be stated precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the BVAE parameter reduction is by construction but the performance claims are externally benchmarked, and the BELBO algebra is the standard ELBO with tied weights.

full rationale

The paper's central empirical claim—that BVAEs 'slightly outperformed' unidirectional VAEs—is tested against standard VAE, beta-VAE, beta-TCVAE, and IWAE baselines implemented with the Pythae framework on four external datasets. No parameter is fitted to a subset of results and then relabeled as a prediction; the reported tables compare independent runs of distinct architectures. The roughly 50% parameter reduction follows by construction from using one network instead of two, but that is an architectural definition, not a loaded prediction. The BELBO derivation in Section III reproduces the standard ELBO inequality: after taking the expectation with respect to the forward-pass density qf(z|x,theta), Eq. (38) is exactly E_{qf}[ln p(x|z,theta)] - KL(qf||p(z)), with the nonnegative KL(qf||q) gap discarded. The paper then estimates this objective using the backward-pass density pb(x|z,theta) in Algorithm 1; if pb is taken to define the generative likelihood p(x|z,theta), the bound is the ordinary ELBO for the BVAE model. The mismatch between the notation p(x|z,theta) in the derivation and pb(x|z,theta) in the algorithm is a correctness or bookkeeping concern, not a circular reduction to the paper's own inputs. The B-BP algorithm and G-NoVa activations are cited from the authors' prior work, but Algorithm 1 spells out the training procedure, and the empirical comparisons stand independently of those citations. No uniqueness theorem or self-citation chain is used to forbid alternatives or force the paper's conclusions. Therefore no circular step is present under the specified criteria.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its constructs are architectural: a shared-weight network and the BELBO objective. The load-bearing assumptions are the unstated equality between the backward-pass decoder and the generative likelihood, the representational adequacy of one shared network, and the validity of the cited B-BP algorithm.

free parameters (3)
  • G-NoVa activation parameters alpha and beta = not reported
    Equation (40) defines a(x) = alpha*x + x*sigma(beta*x) with alpha > 0 and beta > 0. The values are never specified, yet they affect every layer's behavior and the reported results.
  • Latent dimension per dataset = 64 (MNIST, Fashion-MNIST), 256 (CIFAR-10, CelebA)
    Chosen by the authors and central to reconstruction, generation, and classification quality. No ablation or sensitivity analysis is provided.
  • Beta hyperparameters in beta-VAE and beta-TCVAE baselines = beta = 0.5, 1.5 for MNIST/Fashion/CIFAR; beta = 0.1 for CelebA
    Hand-chosen baseline hyperparameters. They are not part of the BVAE itself but affect the comparison tables.
assumptions (5)
  • domain assumption The intractable posterior q(z|x,theta) in Equation (4) is a true conditional density, so p(x|theta) = p(x,z|theta)/q(z|x,theta) holds pointwise.
    Equations (4) and (26). This holds only if q is the true posterior. The paper never states that the backward-pass decoder pb equals the p(x|z,theta) appearing in the identity.
  • domain assumption A single network with shared weights theta can represent both the encoder qf(z|x,theta) and the decoder pb(x|z,theta) well enough to model the data.
    Section III and Algorithm 1. No capacity analysis, consistency condition, or empirical verification that the shared representation is sufficient.
  • domain assumption Bidirectional backpropagation from prior work [2] provides valid gradients for the shared-weight objective.
    The paper cites [2] without deriving the algorithm. The central optimization depends entirely on this unstated prior method.
  • domain assumption The decoder output loss (BCE) corresponds to a valid log-likelihood ln pb(x|z,theta).
    Algorithm 1 uses BCE without specifying the output distribution (Bernoulli vs Gaussian) or its parameters, so the likelihood interpretation is incomplete.
  • standard math Jensen's inequality and KL nonnegativity as used in the ELBO derivation.
    Equations (18)-(22) rely on these standard results, which are valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bidirectional Variational Autoencoders." pith.science (2026). https://pith.science/paper/WZM3ROB3

@misc{pith2026250516074,
  author       = {Pith},
  title        = {Pith review of: Bidirectional Variational Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZM3ROB3}},
  note         = {Machine review of arXiv:2505.16074}
}
read the original abstract

We present the new bidirectional variational autoencoder (BVAE) network architecture. The BVAE uses a single neural network both to encode and decode instead of an encoder-decoder network pair. The network encodes in the forward direction and decodes in the backward direction through the same synaptic web. Simulations compared BVAEs and ordinary VAEs on the four image tasks of image reconstruction, classification, interpolation, and generation. The image datasets included MNIST handwritten digits, Fashion-MNIST, CIFAR-10, and CelebA-64 face images. The bidirectional structure of BVAEs cut the parameter count by almost 50% and still slightly outperformed the unidirectional VAEs.

Figures

Figures reproduced from arXiv: 2505.16074 by the authors.

Figure 1
Figure 1. Bidirectional vs. unidirectional variational autoencoders: Unidirectional VAEs use the forward passes of two separate networks for encoding and decoding. Bidirectional VAEs encode on the forward pass and decode on the backward pass with the same synaptic weight matrices in both directions. This cuts the number of tunable parameters roughly in half. (a) The decoder network with parameter θ approximates p(x|z, θ) and … view at source ↗
Figure 2
Figure 2. BELBO training of a bidirectional variational autoencoder with the bidirectional backpropagation algorithm. BELBO maximization uses a single network for encoding and decoding. The forward pass with likelihood qf (z|x, θ) encodes the latent features. The backward pass with likelihood pb(x|z, θ) decodes the latent features over the same web of synapses. and DKL qf (z|x, ϕ)||q(z|x, θ)  = Ez|x,ϕ ln qf (z|x, ϕ) q(z|x, … view at source ↗
Figure 3
Figure 3. Bidirectional VAE with residual network architecture: This cuts the tunable parameters roughly in half compared with unidirectional VAEs. (a) is the bidirectional convolutional layer. Convolution runs in the forward pass and convolution transpose runs in reverse with the same set of convolution masks. (b) is the architecture of a bidirectional residual block with bidirectional skip connections. The Fashion-MNIST dat… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: t-SNE embedding for the MNIST handwritten digit dataset: Latent space dimension is 128. (a) A simple linear classifier that trained on the unidirectional VAE-compressed features achieved a 95.2% accuracy. (b) The simple classifier achieved 97.32% accuracy when it train…
Figure 5
Figure 5. Figure 5: MNIST handwritten image: Image interpolation with variational autoencoder networks. TABLE II: Fashion-MNIST dataset with VAEs. We used the residual network architecture. The dimension of the latent variable is 64. The BVAEs use 42.2MB memory parameter and the unidirect…
Figure 6
Figure 6. Figure 6: Image interpolation with VAEs on the Fashion-MNIST dataset. TABLE III: CIFAR-10 dataset with VAEs. The dimension of the latent space is 256. The BVAEs each used 107MB memory parameter and the unidirectional VAEs each used 214MB memory parameter. Model Parameters Genera…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 19 canonical work pages

  1. [2]

    Bidirectional backpropagation,

    O. Adigun and B. Kosko, “Bidirectional backpropagation,” IEEE Trans. Syst. Man Cybern. Syst. , vol. 50, no. 5, pp. 1982–1994, 2020

    work page 1982

  2. [1]

    Bidirectional representation and backpropa- gation learning,

    O. Adigun and B. Kosko, “Bidirectional representation and backpropa- gation learning,” in International Joint Conference on Advances in Big Data Analytics, ABDA 2016 . CSREA, 2016, pp. 3–9

    work page 2016

  3. [3]

    Deeper neural networks with non-vanishing logistic hidden units: Nova vs. relu neurons,

    O. Adigun and B. Kosko, “Deeper neural networks with non-vanishing logistic hidden units: Nova vs. relu neurons,” in 2021 20th IEEE Inter- national Conference on Machine Learning and Applications (ICMLA) . IEEE, 2021, pp. 1407–1412

    work page 2021

  4. [4]

    Deeper bidirectional neural networks with generalized non-vanishing hidden neurons,

    O. Adigun and B. Kosko, “Deeper bidirectional neural networks with generalized non-vanishing hidden neurons,” in 21st IEEE International Conference on Machine Learning and Applications, ICMLA 2022, Nassau, Bahamas, December 12-14, 2022 . IEEE, 2022, pp. 69–76

    work page 2022

  5. [5]

    Bidirectional backpropagation autoencoding networks for image compression and denoising,

    O. Adigun and B. Kosko, “Bidirectional backpropagation autoencoding networks for image compression and denoising,” in 2023 International Conference on Machine Learning and Applications (ICMLA) . IEEE, 2023, pp. 730–737

    work page 2023

  6. [6]

    Importance weighted autoencoders,

    Y . Burda, R. Grosse, and R. Salakhutdinov, “Importance weighted autoencoders,” arXiv preprint arXiv:1509.00519 , 2015

    arXiv 2015

  7. [7]

    Importance weighted autoencoders,

    Y . Burda, R. B. Grosse, and R. Salakhutdinov, “Importance weighted autoencoders,” in 4th International Conference on Learning Represen- tations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016

    work page 2016

  8. [8]

    Understanding disentangling in β-vae,

    C. P. Burgess, I. Higgins, A. Pal, L. Matthey, N. Watters, G. Desjardins, and A. Lerchner, “Understanding disentangling in β-vae,” CoRR, vol. abs/1804.03599, 2018. [Online]. Available: http://arxiv.org/abs/1804.03599

    arXiv 2018

Show all 34 references
  1. [9]

    Hamiltonian variational auto-encoder,

    A. L. Caterini, A. Doucet, and D. Sejdinovic, “Hamiltonian variational auto-encoder,” in Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montr´eal, Canada, 2018, pp. 8178– 8188

    2018

  2. [10]

    Pythae: Unifying generative autoencoders in python - A benchmarking use case,

    C. Chadebec, L. J. Vincent, and S. Allassonni `ere, “Pythae: Unifying generative autoencoders in python - A benchmarking use case,” in NeurIPS, 2022

    2022

  3. [11]

    Isolating sources of disentanglement in variational autoencoders,

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

    2018

  4. [12]

    Hyperspherical variational auto-encoders,

    T. R. Davidson, L. Falorsi, N. D. Cao, T. Kipf, and J. M. Tomczak, “Hyperspherical variational auto-encoders,” inProceedings of the Thirty- Fourth Conference on Uncertainty in Artificial Intelligence, UAI 2018, Monterey, California, USA, August 6-10, 2018 . AUAI Press, 2018, p...

    2018

  5. [13]

    A guide to convolution arithmetic for deep learning,

    V . Dumoulin and F. Visin, “A guide to convolution arithmetic for deep learning,” arXiv preprint arXiv:1603.07285 , 2016

    2016 arXiv

  6. [14]

    Sur la distance de deux lois de probabilit ´e,

    M. Fr ´echet, “Sur la distance de deux lois de probabilit ´e,” in Annales de l’ISUP, vol. 6, no. 3, 1957, pp. 183–198

    1957

  7. [15]

    From variational to deterministic autoencoders,

    P. Ghosh, M. S. M. Sajjadi, A. Vergari, M. J. Black, and B. Sch ¨olkopf, “From variational to deterministic autoencoders,” in 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 , 2020

    2020

  8. [16]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017,...

    2017

  9. [17]

    beta-vae: Learning basic visual concepts with a constrained variational framework,

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” in International conference on learning representations , 2016

    2016

  10. [18]

    Reducing the dimensionality of data with neural networks,

    G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” science, vol. 313, no. 5786, pp. 504–507, 2006

    2006

  11. [19]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings , 2014

    2014

  12. [20]

    Bidirectional associative memories,

    B. Kosko, “Bidirectional associative memories,” IEEE Transactions on Systems, Man, and Cybernetics , vol. 18, no. 1, pp. 49–60, 1988

    1988

  13. [21]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hinton et al. , “Learning multiple layers of features from tiny images,” 2009

    2009

  14. [22]

    Classification using discriminative re- stricted boltzmann machines,

    H. Larochelle and Y . Bengio, “Classification using discriminative re- stricted boltzmann machines,” in Proceedings of the 25th international conference on Machine learning , 2008, pp. 536–543

    2008

  15. [23]

    The mnist database of handwritten digits,

    Y . LeCun, “The mnist database of handwritten digits,” http://yann. lecun. com/exdb/mnist/, 1998

    1998

  16. [24]

    Deep learning face attributes in the wild,

    Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in Proceedings of International Conference on Computer Vision (ICCV), December 2015

    2015

  17. [25]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

    2017 arXiv

  18. [26]

    Tighter variational bounds are not necessarily better,

    T. Rainforth, A. Kosiorek, T. A. Le, C. Maddison, M. Igl, F. Wood, and Y . W. Teh, “Tighter variational bounds are not necessarily better,” in International Conference on Machine Learning . PMLR, 2018, pp. 4277–4285

    2018

  19. [27]

    Consistency regularization for variational auto-encoders,

    S. Sinha and A. B. Dieng, “Consistency regularization for variational auto-encoders,” in Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual , 2021, pp. 12 943–12 954

    2021

  20. [28]

    Super-convergence: Very fast training of neural networks using large learning rates,

    L. N. Smith and N. Topin, “Super-convergence: Very fast training of neural networks using large learning rates,” in Artificial intelligence and machine learning for multi-domain operations applications , vol. 11006. SPIE, 2019, pp. 369–386

    2019

  21. [29]

    Information processing in dynamical systems: Foundations of harmony theory,

    P. Smolensky et al. , “Information processing in dynamical systems: Foundations of harmony theory,” 1986

    1986

  22. [30]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research , vol. 9, no. 11, 2008

    2008

  23. [31]

    Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion,

    P. Vincent, H. Larochelle, I. Lajoie, Y . Bengio, and P.-A. Manzagol, “Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion,” Journal of Machine Learning Research , vol. 11, no. 110, pp. 3371–3408, 2010. [Online]. A...

    2010

  24. [32]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  25. [33]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” arXiv preprint arXiv:1708.07747, 2017

    2017 arXiv

  26. [34]

    Infovae: Information maximizing variational autoencoders,

    S. Zhao, J. Song, and S. Ermon, “Infovae: Information maximizing variational autoencoders,” arXiv preprint arXiv:1706.02262 , 2017

    2017 arXiv

Pith tools

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