Pith. sign in

REVIEW 3 major objections 6 minor 35 references

DGSAN: Discrete Generative Self-Adversarial Network

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read DGSAN trains discrete generative models adversarially by deriving each new generator from the last discriminator, with no gradient passed through discrete samples.

desk verdict DGSAN has a genuinely novel training scheme for discrete GANs, but its convergence proof rests on an unverified pointwise condition, making it a serious paper that needs revision rather than a desk reject. read the letter →

arxiv 1908.09127 v2 pith:LEV45KZ5 submitted 2019-08-24 cs.LG stat.ML

classification cs.LGstat.ML
keywords discretedatagenerationgenerativeadversarialnetworkssequenceself-adversarialtrainingexplicitmodelsJensen-Shannondivergencef-divergenceBregman
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 DGSAN, a way to train a generative model of discrete data (text) adversarially while never passing gradients through the generator's discrete sampling step. Instead of sending a reward or a relaxed gradient to the generator, each iteration treats the current generator $Q_{\text{old}}$ as the fake distribution and derives the next generator $Q_{\theta}$ directly from a discriminator of the form $q_{\theta}/(q_{\theta}+q_{\text{old}})$; the objective uses only samples from the real distribution and from $Q_{\text{old}}$. The authors prove that in the ideal case this drives $Q_{\theta}$ to the real distribution $P$, and that if the learned discriminator stays between a random guess and the optimal discriminator for $Q_{\text{old}}$, the Jensen-Shannon divergence strictly decreases each iteration, so the sequence of generators converges to $P$. The argument extends to any $f$-divergence with strictly convex $f$. On three text datasets, DGSAN reports better semantic Frechet distance and n-gram BLEU scores than maximum likelihood, variational-autoencoder, and several GAN-based baselines, although its negative log-likelihood stays worse than maximum likelihood.

What carries the argument

The self-discriminator identity: for a candidate next generator $Q_{\theta}$ and the previous generator $Q_{\text{old}}$, the discriminator is fixed as $D(x)=q_{\theta}(x)/(q_{\theta}(x)+q_{\text{old}}(x))$, exactly the form the optimal GAN discriminator would take if $Q_{\theta}$ were the data distribution. Substituting this into the GAN objective yields Eq.~5, whose dependence on $\theta$ is only through $q_{\theta}$ evaluated at samples drawn from $P$ and $Q_{\text{old}}$, so no gradient crosses a discrete sample. The theoretical engine is a Bregman-divergence decomposition $D_{\mathrm{JS}}(P\|Q_{\text{old}}) = L(P,Q_{\text{old}},Q_{\theta}) + \mathbb{E}_{Q_{\text{old}}}[B_f(p/q_{\text{old}}\|q_{\theta}/q_{\text{old}})]$ with $f(u)=u\ln u-(u+1)\ln(u+1)$, which shows that maximizing $L$ minimizes the Bregman remainder and estimates the ratio $p/q_{\text{old}}$ by $q_{\theta}/q_{\text{old}}$; an analogous identity holds for every strictly convex $f$ through Fenchel conjugation. For sequences, a curriculum in block length $l$ is added so the model learns $P(x_k..x_{k+l}\mid x_1..x_{k-1})$ and sees generated samples as prefixes grow.

What would settle it

Run Algorithm 1 on a discrete distribution and, at each iteration, test the betweenness condition on a large sample by comparing $D_{\theta}(x)$ with $0.5$ and $D^*(x)=p(x)/(p(x)+q_{\text{old}}(x))$; simultaneously estimate $D_{\mathrm{JS}}(P\|Q_{\theta})-D_{\mathrm{JS}}(P\|Q_{\text{old}})$ on held-out data. One iteration that violates the condition and still shows a nonnegative divergence change refutes the claimed sufficiency of the update; repeated violations with decreasing divergence would show the condition is not necessary, which the paper does not claim.

Watch

Extended reading notes

Core claim

The central discovery is that in finite discrete domains an explicit generative model can be trained adversarially by solving for the next generator in closed form instead of backpropagating through samples. The standard GAN optimal-discriminator relation $D^*(x)=p(x)/(p(x)+q_{\text{old}}(x))$ is inverted: writing the next generator as $q_{\theta}$ and fixing the discriminator as $D(x)=q_{\theta}(x)/(q_{\theta}(x)+q_{\text{old}}(x))$, the objective $L(P,Q_{\text{old}},Q_{\theta})=\mathbb{E}_{x\sim P}[\ln(q_{\theta}/(q_{\theta}+q_{\text{old}}))]+\mathbb{E}_{x\sim Q_{\text{old}}}[\ln(q_{\text{old}}/(q_{\theta}+q_{\text{old}}))]$ is maximized with respect to $\theta$. At the global optimum $Q_{\theta}$ equals $P$; short of that, Theorems 4.2 and 4.4 show that whenever $D_{\theta}$ lies pointwise between $0.5$ and $D^*$, the Jensen-Shannon (or, more generally, $f$-) divergence between the generator and the real distribution strictly shrinks, so the iteration converges. For sequences, the joint distribution is learned by curriculum training over blocks of increasing length conditioned on real prefixes, which the paper argues progressively alleviates exposure bias.

Load-bearing premise

The convergence guarantee holds only if at every iteration the learned self-discriminator $D_{\theta}(x)$ stays pointwise between the random baseline $0.5$ and the optimal discriminator $D^*(x)=p(x)/(p(x)+q_{\text{old}}(x))$ for the current generator; the algorithm neither enforces nor verifies this condition, and if it fails the strict decrease in divergence does not follow.

Editorial extensions

If this is right

  • Discrete GAN training no longer needs policy-gradient estimators, Gumbel-Softmax relaxations, or a separate discriminator network; the same network provides both the next generator and the real-vs-fake decision.
  • If the betweenness condition holds at every iteration, the generator provably moves closer to the real distribution in Jensen-Shannon divergence, so the usual saddle-point instability of GAN training is bypassed.
  • The framework applies to any finite discrete domain with an explicit tractable generative distribution, not only to natural-language sequences.
  • On three text corpora the method reports lower Frechet feature distance and higher BLEU than maximum-likelihood, variational-autoencoder, and GAN baselines, while its negative log-likelihood is worse than maximum likelihood; the experiments also show the training loss settles near $2\ln 2$, the theoretical no-discrimination value.

Reading between the lines

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

  • The betweenness condition of Theorem 4.2 is stated but neither enforced nor verified during training, so the practical convergence guarantee rests on an assumption the algorithm could silently violate; a violated step need not produce a divergence decrease even though the objective was maximized.
  • Because each iteration fits $q_{\theta}/q_{\text{old}}$ to $p/q_{\text{old}}$, DGSAN is effectively an iterative density-ratio estimation scheme; this suggests direct connections to ratio-estimation and boosting algorithms, where the current model serves as the reference measure.
  • The block-length curriculum makes a testable prediction: exposure bias should diminish as the trained block length $l$ grows, and vanish when $l$ reaches the full sequence length, since generated prefixes then appear during training.
  • The same closed-form discriminator could serve as an estimator of the likelihood ratio $p/q_{\text{old}}$ outside generation, for example in divergence-based training objectives or importance-weighting schemes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 DGSAN, a method for training explicit generative models of discrete sequences without backpropagating gradients through discrete samples. At each iteration, a generator Qθ is trained by maximizing the objective in Eq. (5) against the previous generator Qold, using a discriminator defined as qθ/(qθ+qold). The authors prove an exact decomposition of the Jensen–Shannon divergence (Theorem 4.1), a conditional convergence theorem (Theorem 4.2), and an f-divergence extension (Theorems 4.3 and 4.4). Experiments on Amazon, Yelp, and COCO compare DGSAN with MLE, VAE, SeqGAN, MaliGAN, and RankGAN using NLL, BLEU, Backward BLEU, MS-Jaccard, and Fréchet BERT Distance, with DGSAN reporting the best FBD and BLEU scores on all three datasets.

Significance. If the convergence guarantee were established, DGSAN would be a useful addition to the discrete-generation toolbox: it avoids RL or Gumbel-softmax gradient estimators, uses a single network, and the JS-decomposition in Theorem 4.1 is a clean mathematical identity. The paper also presents the softplus objective of Eq. (6), which is exactly equivalent to the proposed objective, and it provides implementation links. However, the main theoretical claim is conditional on an interlacing condition that the algorithm does not enforce or verify, and the tempered sampling in the sequence-generation algorithm deviates from the analyzed objective. The empirical results are mixed across metrics (e.g., NLL is worse than MLE on two of three datasets), and the headline metric FBD comes from the authors' own prior work, so the strength of the empirical claims as stated is not fully supported. With revisions that close the theory–algorithm gap and temper the empirical claims, the contribution could be solid.

major comments (3)
  1. [§4.2, Theorem 4.2, and Appendix Eq. (35)] The convergence guarantee is conditional on an assumption that Algorithm 1 neither enforces nor checks. The proof of Theorem 4.4 shows that the premise 'Dθ is between a random and an optimal discriminator' is equivalent to qθ(x) lying pointwise between qold(x) and p(x) for every x (Eq. 35). Algorithm 1 optimizes Eq. (5) by unconstrained gradient descent on θ, with no constraint of this interlacing form and no verification that the optimized qθ satisfies it. Moreover, the unconstrained maximizer of Eq. (5) is qθ = p, which is an endpoint of the interval rather than a point strictly inside it, so the strict inequality in the premise can fail even in the ideal case. Since the authors state that reaching the global optimum 'is ideal but may be impossible' (Section 4.2), Theorem 4.2 is the only support for the non-global regime; as written, the claimed per-iteration strict decrease of DJS and convergence to P is not established for the implemented algorithm.
  2. [§4.3 and Algorithm 2, line 10] The sequence-generation algorithm samples the 'fake' continuations from Qold with a sampling temperature T = 2.0, whereas the objective in Eq. (11) and the theory in Section 4.2 are stated for samples from Qold itself. Sampling from a tempered distribution is not the same as sampling from Qold for T ≠ 1, so the empirical expectation in line 10 is not an unbiased estimate of the second term in Eq. (11). The paper does not discuss this mismatch or modify the theory to account for temperature; this is a second, distinct gap between the implementation and the convergence analysis.
  3. [§5.3 and Tables 1–3] The empirical comparison is built around FBD, a metric introduced by the authors in [27], and the same paper's MS-Jaccard is also reported; all checkpoints are selected by FBD. This does not by itself invalidate the comparison across methods, since the same selection rule is applied to all models, but it makes the headline metric non-independent, and no standard errors or significance tests are reported. More importantly, the claim of 'generally the superiority' is not consistently supported by the tables: DGSAN has worse NLL than MLE on Amazon (113.146 vs 99.369) and Yelp (56.398 vs 50.201), and worse MSJ3 than MLE on Amazon (0.442 vs 0.560) and COCO (0.208 vs 0.224). The conclusions should be rephrased to match the mixed evidence.
minor comments (6)
  1. [Appendix, proof of Theorem 4.2] The proof refers to 'Theorem 3.2' where it should refer to Theorem 4.2.
  2. [§4.2, Theorem 4.2] The term 'random discriminator' is used in the theorem statement but is never defined; from the proof it is evidently the constant 0.5, and this should be stated explicitly.
  3. [§4.3, Algorithm 2] The loop variable l is incremented on every outer iteration, but the termination condition is 'until max epoch'; the relationship between l and the maximum sequence length M, and the intended curriculum schedule, should be clarified in the pseudocode.
  4. [§5.3] The hyperparameters D (number of DGSAN iterations per length) and T (sampling temperature) are fixed without sensitivity analysis; given that temperature changes the sampling distribution, a small study of T would be valuable.
  5. [§5.4, Figure 1] The caption should state which loss is plotted (presumably the softplus objective of Eq. (6)) and how the claimed theoretical value 2 ln 2 ≈ 1.38 follows for that loss.
  6. [§2 and §5] The related work discusses TextGAN, FM-GAN, LeakGAN, and RelGAN, but none of these methods appear in the experimental comparison; the claim of competitiveness with 'popular or recent methods' is therefore narrower than the text suggests.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DGSAN's generator update is a variational identity, and the only self-referential element (author-defined FBD/MSJ metrics) is not load-bearing.

full rationale

The derivation chain is self-contained. Equation 5 defines the objective L(P, Qold, Qθ), whose pointwise optimum over qθ is qθ = p, and Theorem 4.1 decomposes the fixed quantity D_JS(P||Qold) into L plus a nonnegative Bregman divergence term, so maximizing L is equivalent to minimizing the Bregman divergence between p/qold and qθ/qold (Corollary 4.1.1). This is a mathematical identity, not a fitted parameter renamed as a prediction. Theorem 4.2/4.4 is explicitly conditional: it assumes Dθ lies pointwise between 0.5 and D*, equivalently qθ lies between qold and p (Appendix Eq. 35), and then derives a strict decrease of the f-divergence. The paper itself states that reaching the global optimum is 'ideal but may be impossible,' and the algorithm neither enforces nor verifies the interlacing condition; this is an unproven premise affecting the convergence guarantee, but the theorem does not assume its own conclusion and no equation reduces to its input by construction. The experimental section cites reference [27], a prior paper by the same authors, for FBD and MS-Jaccard, and Section 5.3 states that all evaluations are based on best checkpoints selected by FBD; this is a self-citation in the evaluation pipeline, but it is not load-bearing for the method's derivation or for the theoretical claims. The method is also compared against external baselines and external metrics (BLEU, Backward BLEU, NLL), so the empirical validation is not self-confirming. Hence no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework rests on a small set of modeling choices: the self-adversarial discriminator ratio, the softplus surrogate loss, and a convergence condition that is not enforced. No new physical entities are postulated.

free parameters (3)
  • Temperature T = 2.0
    Used in Algorithm 2 when sampling from Qold; affects the training distribution and is chosen without sensitivity analysis.
  • DGSAN iterations per length D = 5
    Number of self-adversarial iterations per sequence length; chosen heuristically.
  • Training epochs = 1000 for DGSAN/MLE/VAE, 280 for GANs
    Different epoch counts may affect comparison; no analysis of convergence or overfitting is provided.
assumptions (4)
  • ad hoc to paper The self-discriminator D(x)=qθ(x)/(qθ(x)+qold(x)) is a valid stand-in for the optimal discriminator p/(p+qold) even when qθ is not optimal.
    Introduced in Section 4.1, Eq. 3; motivated by the fixed-point property but not derived from first principles.
  • domain assumption The optimization objective Eq. 5 (or its softplus surrogate Eq. 6) can be globally optimized to reach qθ=P.
    The theory in Section 4.2 assumes global optimality of L; in practice, non-convex neural network optimization may not reach it.
  • ad hoc to paper The condition of Theorem 4.2 (Dθ between random and optimal) holds at each iteration of Algorithm 1.
    The proof of convergence depends on this condition, but the algorithm does not guarantee it.
  • domain assumption The explicit generative model qθ(x) can be evaluated and differentiated tractably (softmax over a finite vocabulary).
    This holds for finite discrete sequence models, as stated in Section 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DGSAN: Discrete Generative Self-Adversarial Network." pith.science (2026). https://pith.science/paper/LEV45KZ5

@misc{pith2026190809127,
  author       = {Pith},
  title        = {Pith review of: DGSAN: Discrete Generative Self-Adversarial Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LEV45KZ5}},
  note         = {Machine review of arXiv:1908.09127}
}
read the original abstract

Although GAN-based methods have received many achievements in the last few years, they have not been entirelysuccessful in generating discrete data. The most crucial challenge of these methods is the difficulty of passing the gradientfrom the discriminator to the generator when the generator outputs are discrete. Despite the fact that several attemptshave been made to alleviate this problem, none of the existing GAN-based methods have improved the performance oftext generation compared with the maximum likelihood approach in terms of both the quality and the diversity. In thispaper, we proposed a new framework for generating discrete data by an adversarial approach in which there is no need topass the gradient to the generator. The proposed method has an iterative manner in which each new generator is definedbased on the last discriminator. It leverages the discreteness of data and the last discriminator to model the real datadistribution implicitly. Moreover, the method is supported with theoretical guarantees, and experimental results generallyshow the superiority of the proposed DGSAN method compared to the other popular or recent methods in generatingdiscrete sequential data.

Figures

Figures reproduced from arXiv: 1908.09127 by the authors.

Figure 1
Figure 1. DGSAN training loss of different datasets [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 21 canonical work pages

  1. [27]

    Alihosseini, E

    D. Alihosseini, E. Montahaei, M. Soleymani Baghshah, i n: Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation, Association for Co mputa- tional Linguistics, Minneapolis, Minnesota, 2019, pp. 90– 98. URL: https://www.aclweb.org/anthology/W19-2311. doi:10.18653/v1/W19-2311

  2. [1]

    Huszar, CoRR abs/1511.05101 (2015)

    F. Huszar, CoRR abs/1511.05101 (2015). arXiv:1511.05101

  3. [2]

    Bengio, O

    S. Bengio, O. Vinyals, N. Jaitly, N. Shazeer, in: C. Corte s, N. D. Lawrence, D. D. Lee, M. Sugiyama, R. Garnett (Eds.), Advance s in Neural Information Processing Systems 28: Annual Confer ence on Neural Information Processing Systems 2015, December 7- 12, 2015, Montreal, Quebec, Canada, pp. 1171–1179

  4. [3]

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Wa rde- Farley, S. Ozair, A. C. Courville, Y . Bengio, CoRR abs/1406. 2661 (2014). arXiv:1406.2661

  5. [4]

    L. Y u, W. Zhang, J. Wang, Y . Y u, in: S. P . Singh, S. Markovit ch (Eds.), Proceedings of the Thirty-First AAAI Conference on Arti- ficial Intelligence, February 4-9, 2017, San Francisco, Cal ifornia, USA., AAAI Press, 2017, pp. 2852–2858

  6. [5]

    I. J. Goodfellow, CoRR abs/1701.00160 (2017). arXiv:1701.00160

  7. [6]

    J. Guo, S. Lu, H. Cai, W. Zhang, Y . Y u, J. Wang, in: S. A. McIlraith, K. Q. Weinberger (Eds.), Proceedings of the Thir ty- Second AAAI Conference on Artificial Intelligence, New Orle ans, Louisiana, USA, February 2-7, 2018, AAAI Press, 2018

  8. [7]

    K. Lin, D. Li, X. He, M. Sun, Z. Zhang, in: [ 33], pp. 3158–3168

Show all 35 references
  1. [8]

    T. Che, Y . Li, R. Zhang, R. D. Hjelm, W. Li, Y . Song, Y . Bengi o, CoRR abs/1702.07983 (2017). arXiv:1702.07983

  2. [9]

    M. J. Kusner, J. M. Hernández-Lobato (2016). URL: https://arxiv.org/abs/1611.04051

  3. [10]

    Zhang, Z

    Y . Zhang, Z. Gan, K. Fan, Z. Chen, R. Henao, D. Shen, L. Car in, in: D. Precup, Y . W. Teh (Eds.), Proceedings of the 34th Internat ional Conference on Machine Learning, ICML 2017, Sydney, NSW, Aus- tralia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Resear...

  4. [11]

    L. Chen, S. Dai, C. Tao, D. Shen, Z. Gan, H. Zhang, Y . Zhang , L. Carin (2018). URL: DBLP:conf/nips/2018

  5. [12]

    W. Nie, N. Narodytska, A. Patel, International Confere nce on Learning Representations (ICLR) (2019)

  6. [13]

    R. D. Hjelm, A. P . Jacob, A. Trischler, G. Che, K. Cho, Y . B engio, in: International Conference on Learning Representations

  7. [14]

    D. P . Kingma, M. Welling, in: Y . Bengio, Y . LeCun (Eds.), 2nd International Conference on Learning Representations, IC LR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proc eed- ings. URL: http://arxiv.org/abs/1312.6114

  8. [15]

    S. R. Bowman, L. Vilnis, O. Vinyals, A. M. Dai, R. Józe- fowicz, S. Bengio, in: Y . Goldberg, S. Riezler (Eds.), Proceedings of the 20th SIGNLL Conference on Compu- tational Natural Language Learning, CoNLL 2016, Berlin, Germany, August 11-12, 2016, ACL, 2016, pp. 10–21. URL: ...

  9. [16]

    X. Chen, D. P . Kingma, T. Salimans, Y . Duan, P . Dhari- wal, J. Schulman, I. Sutskever, P . Abbeel, in: 5th In- ternational Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Confer- ence Track Proceedings, OpenReview.net, 2017. URL: https:...

  10. [17]

    S. Zhao, J. Song, S. Ermon, in: The Thirty-Third AAAI Con ference on Artificial Intelligence, AAAI 2019, The Thirty-First Inn ovative Applications of Artificial Intelligence Conference, IAAI 2 019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI ...

  11. [18]

    J. He, D. Spokoyny, G. Neubig, T. Berg-Kirkpatrick, in: 7th Interna- tional Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, OpenReview.net, 2019. URL: https://openreview.net/forum?id=rylDfnCqF7

  12. [19]

    Dugas, Y

    C. Dugas, Y . Bengio, F. Bélisle, C. Nadeau, R. Garcia, in: T. K. Leen, T. G. Dietterich, V . Tresp (Eds.), Ad- vances in Neural Information Processing Systems 13, Papers from Neural Information Processing Systems (NIPS) 2000, Denver, CO, USA, MIT Press, 2000, pp. 472–478. URL...

  13. [20]

    Sugiyama, T

    M. Sugiyama, T. Suzuki, T. Kanamori, Annals of the Insti tute of Statistical Mathematics 64 (2012) 1009–1044

  14. [21]

    Nowozin, B

    S. Nowozin, B. Cseke, R. Tomioka, in: D. D. Lee, M. Sugiya ma, U. von Luxburg, I. Guyon, R. Garnett (Eds.), Advances in Neur al Information Processing Systems 29: Annual Conference on Ne u- ral Information Processing Systems 2016, December 5-10, 20 16, Barcelona, Spain, pp. 271–279

  15. [22]

    Poole, A

    B. Poole, A. A. Alemi, J. Sohl-Dickstein, A. Angelova, C oRR abs/1612.02780 (2016). arXiv:1612.02780

  16. [23]

    Caccia, L

    M. Caccia, L. Caccia, W. Fedus, H. Larochelle, J. Pineau, L. Charlin, in: International Confer- ence on Learning Representations (ICLR). URL: https://openreview.net/forum?id=BJgza6VtPB

  17. [24]

    Papineni, S

    K. Papineni, S. Roukos, T. Ward, W. Zhu, in: Proceedings of the 40th Annual Meeting of the Association for Computational Li n- guistics, July 6-12, 2002, Philadelphia, PA, USA., ACL, 200 2, pp. 311–318

  18. [25]

    Y . Zhu, S. Lu, L. Zheng, J. Guo, W. Zhang, J. Wang, Y . Y u, in: K. Collins-Thompson, Q. Mei, B. D. Davison, Y . Liu, E. Yil - maz (Eds.), The 41st International ACM SIGIR Conference on R e- search & Development in Information Retrieval, SIGIR 2018, Ann Arbor, MI, USA, July 0...

  19. [26]

    Zhang, V

    T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, Y . Artzi, C oRR abs/1904.09675 (2019). arXiv:1904.09675

  20. [28]

    Z. Shi, X. Chen, X. Qiu, X. Huang, Proceed- ings of the Twenty-Seventh International Joint Con- ference on Artificial Intelligence (2018). URL: http://dx.doi.org/10.24963/ijcai.2018/606. doi:10.24963/ijcai.2018/606

  21. [29]

    Heusel, H

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, S. H ochreiter, in: [ 33], pp. 6629–6640

  22. [30]

    Devlin, M

    J. Devlin, M. Chang, K. Lee, K. Toutanova, CoRR abs/1810 .04805 (2018). arXiv:1810.04805

  23. [31]

    T. Shen, T. Lei, R. Barzilay, T. S. Jaakkola, in: I. Guyon , U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V . N. Vishwanathan, R. Garnett (Eds.), Advances in Neu- ral Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017,...

  24. [32]

    J. J. McAuley, J. Leskovec, in: Q. Y ang, I. King, Q. Li, P . Pu, G. Karypis (Eds.), Seventh ACM Conference on Recommender Systems, RecSys ’13, Hong Kong, China, October 12-16, 2013, ACM, 2013, pp. 165–172. URL: https://doi.org/10.1145/2507157.2507163. doi:10.1145/2507157.2507163

  25. [33]

    Guyon, U

    I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fer gus, S. V . N. Vishwanathan, R. Garnett (Eds.), Advances in Neural In- formation Processing Systems 30: Annual Conference on Neur al Information Processing Systems 2017, 4-9 December 2017, Lo ng Beach, CA, USA, 2017. 9

  26. [34]

    Lemma 7.1

    Appendix Proofs of the theorems presented in Section 4.2 are de- scribed here. Lemma 7.1. If f is convex then f ∗( f ′(x)) = f ′(x)x − f (x), (12) where f ∗ is the Fenchel conjugate of f . Proof. For each convex function f : f (u) ≥ f ′(x)(u − x) + f (x) ⇒ x f ′(x) − f (x) ≥ u...

  27. [35]

    lobster

    Samples of generated sentences Here are the generated samples of different models on the Image Coco captions, Y elp Restaurant reviews, and Amazon datasets. In addition to reporting the randomly generated samples of all models, we sorted samples of dif- ferent models based on ...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.