pith. sign in

arxiv: 1906.11080 · v1 · pith:CML2LV6Fnew · submitted 2019-06-25 · 💻 cs.LG · cs.AI· stat.ML

AGAN: Towards Automated Design of Generative Adversarial Networks

Hanchao Wang , Jun Huan This is my paper

Pith reviewed 2026-05-25 16:31 UTC · model grok-4.3

classification 💻 cs.LG cs.AIstat.ML
keywords neural architecture searchgenerative adversarial networksGAN designautomated machine learningimage generationCIFAR-10transferable modules
0
0 comments X

The pith

A neural architecture search algorithm finds GAN architectures that outperform human-designed models on unsupervised image generation.

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

The paper introduces AGAN as the first neural architecture search method built specifically for generative adversarial networks rather than borrowing from classification models. It reports that the search locates architectures for unsupervised generation on CIFAR-10 that exceed state-of-the-art performance when the same regularization is applied. The same procedure yields competitive results on supervised tasks at 32 by 32 resolution and produces modules that transfer to STL-10. A reader would care because GAN architecture has previously required expert trial-and-error, so an automated method could accelerate progress in generative modeling without repeated human intervention.

Core claim

We present the first neural architecture search algorithm, automated neural architecture search for deep generative models, or AGAN, that is specifically suited for GAN training. For unsupervised image generation tasks on CIFAR-10, our algorithm finds architecture that outperforms state-of-the-art models under same regularization techniques. For supervised tasks, the automatically searched architectures also achieve highly competitive performance, outperforming best human-invented architectures at resolution 32×32. Moreover, we empirically demonstrate that the modules learned by AGAN are transferable to other image generation tasks such as STL-10.

What carries the argument

AGAN, the automated neural architecture search algorithm tailored for GAN training that explores and evaluates candidate generator and discriminator architectures

If this is right

  • GAN architecture design can shift from manual trial-and-error to an automated search process.
  • Unsupervised image generation on CIFAR-10 reaches higher performance without changes to regularization.
  • Architectures discovered at 32 by 32 resolution remain competitive for supervised tasks.
  • Modules found on one dataset transfer directly to image generation on STL-10.
  • Architectural improvements become a scalable route to better GAN training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach may generalize to search spaces that include higher-resolution or conditional generation tasks.
  • Transfer of modules suggests partial reuse could reduce compute when moving between datasets.
  • If the search space is representative, many existing human GAN designs may sit below the achievable frontier.
  • Similar automated search could be applied to other generative model families once the evaluation protocol is adapted.

Load-bearing premise

The search procedure can reach architectures superior to human designs and the performance evaluations used to guide and compare it remain unbiased.

What would settle it

A controlled experiment that applies the identical search constraints, training protocol, and evaluation metric to a top human-designed GAN and finds equal or better performance on CIFAR-10 would falsify the superiority claim.

Figures

Figures reproduced from arXiv: 1906.11080 by Hanchao Wang, Jun Huan.

Figure 1
Figure 1. Figure 1: Controller RNN architecture. Above: The controller consists of three segments, programming the up-sampling [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An normal module defined by controller sequence [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Meta-architecture of the generator and discriminator [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Progression of Inception Score on CIFAR-10 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Images generated by AGANs in supervised image generations tasks [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Topology of modules in AGAN-A Note that the up-sample (down-sample) operations following prev will only be applied when the module is preceded by an up-sampling (down-sampling) module. Our STL-10 network has the same meta-architecture as the one for CIFAR-10, with the distinction that the first up-sampling module in G takes input size of 6 × 6 × n (instead of 4 × 4 × n). We resize the STL-10 data set to 48… view at source ↗
Figure 7
Figure 7. Figure 7: Empirical distribution of sampled operations by module over time [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
read the original abstract

Recent progress in Generative Adversarial Networks (GANs) has shown promising signs of improving GAN training via architectural change. Despite some early success, at present the design of GAN architectures requires human expertise, laborious trial-and-error testings, and often draws inspiration from its image classification counterpart. In the current paper, we present the first neural architecture search algorithm, automated neural architecture search for deep generative models, or AGAN for abbreviation, that is specifically suited for GAN training. For unsupervised image generation tasks on CIFAR-10, our algorithm finds architecture that outperforms state-of-the-art models under same regularization techniques. For supervised tasks, the automatically searched architectures also achieve highly competitive performance, outperforming best human-invented architectures at resolution $32\times32$. Moreover, we empirically demonstrate that the modules learned by AGAN are transferable to other image generation tasks such as STL-10.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces AGAN, the first neural architecture search (NAS) algorithm designed specifically for GANs. It claims that AGAN discovers generator/discriminator architectures outperforming state-of-the-art human-designed models on unsupervised CIFAR-10 image generation under identical regularization, achieves competitive results on supervised tasks at 32x32 resolution, and yields transferable modules to STL-10.

Significance. If the empirical claims hold under rigorous statistical controls, the result would be significant for automating GAN design, reducing reliance on human expertise and trial-and-error. The work provides the first dedicated NAS framework for generative models and demonstrates transferability, which could accelerate progress in unsupervised image synthesis.

major comments (2)
  1. [§4.2, Table 2] §4.2 and Table 2: The CIFAR-10 unsupervised results report single-run FID and IS values for the AGAN-discovered architecture without means or standard deviations across independent training runs (or at least 3–5 seeds). Given the well-known high variance of GAN training and the selection bias inherent in evaluating hundreds of candidates during search, this undermines the central claim of verifiable outperformance over baselines such as SN-GAN and BigGAN under the same regularization.
  2. [§3.2] §3.2 (search space and evaluation protocol): The paper does not specify whether the final reported architecture was re-trained from scratch after search or whether its score was taken from the search-phase evaluation; if the latter, the reported gains may reflect overfitting to the search validation split rather than true architectural superiority.
minor comments (2)
  1. [§4.1] §4.1: The description of the controller and reward function lacks explicit pseudocode or a clear equation for how the reinforcement-learning objective is computed; adding this would improve reproducibility.
  2. [§4.3, Figure 3] Figure 3 and §4.3: The STL-10 transfer experiment does not state whether the transferred modules were frozen or fine-tuned, nor the exact number of epochs used; this detail is needed to interpret the reported gains.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point by point below, indicating planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§4.2, Table 2] §4.2 and Table 2: The CIFAR-10 unsupervised results report single-run FID and IS values for the AGAN-discovered architecture without means or standard deviations across independent training runs (or at least 3–5 seeds). Given the well-known high variance of GAN training and the selection bias inherent in evaluating hundreds of candidates during search, this undermines the central claim of verifiable outperformance over baselines such as SN-GAN and BigGAN under the same regularization.

    Authors: We agree that single-run reporting is insufficient to fully substantiate the claims given GAN training variance and the search process. In the revised manuscript we will add results from at least three independent training runs with different random seeds for the final AGAN architecture (and, where feasible, for the main baselines under identical regularization) and report means together with standard deviations for both FID and IS. revision: yes

  2. Referee: [§3.2] §3.2 (search space and evaluation protocol): The paper does not specify whether the final reported architecture was re-trained from scratch after search or whether its score was taken from the search-phase evaluation; if the latter, the reported gains may reflect overfitting to the search validation split rather than true architectural superiority.

    Authors: The reported performance was obtained by re-training the discovered architecture from scratch on the full training set after the search concluded; the search-phase evaluations were used only to rank candidate architectures. We will add an explicit statement of this protocol in Section 3.2 of the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical NAS results are self-contained experiments

full rationale

The paper presents AGAN as a neural architecture search procedure whose central claims consist of reported performance numbers obtained by executing the search on CIFAR-10 and STL-10. No mathematical derivation, uniqueness theorem, fitted parameter renamed as prediction, or self-citation chain is invoked to establish the superiority result; the outperformance is an observed experimental outcome rather than a quantity forced by construction from the inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, providing no specific information on free parameters, axioms, or invented entities used in the work.

pith-pipeline@v0.9.0 · 5678 in / 1124 out tokens · 72729 ms · 2026-05-25T16:31:09.204146+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive

    cs.CL 2024-02 conditional novelty 6.0

    DPOP is a new loss function that prevents DPO from lowering preferred response likelihoods and outperforms standard DPO on diverse datasets, MT-Bench, and enables Smaug-72B to exceed 80% on the Open LLM Leaderboard.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · cited by 1 Pith paper · 33 internal anchors

  1. [1]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 2672–2680. Curran Associates, Inc., 2014

    work page 2014

  2. [2]

    Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

    Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. CoRR, abs/1609.04802, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  3. [3]

    Generative Adversarial Text to Image Synthesis

    Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. CoRR, abs/1605.05396, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  4. [4]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. CoRR, abs/1703.10593, 2017

    work page arXiv 2017

  5. [5]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning , volume 70 of Proceedings of Machine Learning Research , pages 214–223, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR

    work page 2017

  6. [6]

    Xudong Mao, Qing Li, Haoran Xie, Raymond Y . K. Lau, and Zhen Wang. Multi-class generative adversarial networks with the L2 loss function. CoRR, abs/1611.04076, 2016. 10 A PREPRINT - J UNE 27, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  7. [7]

    f-gan: Training generative neural samplers using variational divergence minimization

    Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. f-gan: Training generative neural samplers using variational divergence minimization. In D. D. Lee, M. Sugiyama, U. V . Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 271–279. Curran Associates, Inc., 2016

    work page 2016

  8. [8]

    Augustus Odena, Jacob Buckman, Catherine Olsson, Tom Brown, Christopher Olah, Colin Raffel, and Ian Goodfellow. Is generator conditioning causally related to GAN performance? In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 3849–38...

    work page 2018

  9. [9]

    Improved Training of Wasserstein GANs

    Ishaan Gulrajani, Faruk Ahmed, Martín Arjovsky, Vincent Dumoulin, and Aaron C. Courville. Improved training of wasserstein gans. CoRR, abs/1704.00028, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  10. [10]

    Spectral Normalization for Generative Adversarial Networks

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. CoRR, abs/1802.05957, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  11. [11]

    Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

    Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. CoRR, abs/1511.06434, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  12. [12]

    Barret Zoph and Quoc V . Le. Neural architecture search with reinforcement learning. CoRR, abs/1611.01578, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  13. [13]

    Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V . Le. Learning transferable architectures for scalable image recognition. CoRR, abs/1707.07012, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  14. [14]

    Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V . Le. Regularized evolution for image classifier architecture search. CoRR, abs/1802.01548, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    Self-Attention Generative Adversarial Networks

    Han Zhang, Ian J. Goodfellow, Dimitris N. Metaxas, and Augustus Odena. Self-attention generative adversarial networks. arXiv:1805.08318, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  16. [16]

    Large Scale GAN Training for High Fidelity Natural Image Synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. CoRR, abs/1809.11096, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  17. [17]

    Conditional Generative Adversarial Nets

    Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. CoRR, abs/1411.1784, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  18. [18]

    cGANs with Projection Discriminator

    Takeru Miyato and Masanori Koyama. cgans with projection discriminator. CoRR, abs/1802.05637, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  19. [19]

    Modulating early visual processing by language

    Harm de Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C. Courville. Modulating early visual processing by language. CoRR, abs/1707.00683, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  20. [20]

    A note on the evaluation of generative models

    Lucas Theis, Aäron van den Oord, and Matthias Bethge. A note on the evaluation of generative models. CoRR, abs/1511.01844, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  21. [21]

    Improved Techniques for Training GANs

    Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. CoRR, abs/1606.03498, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  22. [22]

    Going Deeper with Convolutions

    Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. CoRR, abs/1409.4842, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  23. [23]

    Efficient architecture search by network transformation

    Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang. Efficient architecture search by network transformation. In AAAI, 2018

    work page 2018

  24. [24]

    Path-Level Network Transformation for Efficient Architecture Search

    Han Cai, Jiacheng Yang, Weinan Zhang, Song Han, and Yong Yu. Path-level network transformation for efficient architecture search. arXiv preprint arXiv:1806.02639, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  25. [25]

    Efficient Neural Architecture Search via Parameter Sharing

    Hieu Pham, Melody Y . Guan, Barret Zoph, Quoc V . Le, and Jeff Dean. Efficient neural architecture search via parameter sharing. CoRR, abs/1802.03268, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  26. [26]

    Designing Neural Network Architectures using Reinforcement Learning

    Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. Designing neural network architectures using reinforcement learning. CoRR, abs/1611.02167, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  27. [27]

    BlockQNN: Efficient Block-wise Neural Network Architecture Generation

    Zhao Zhong, Zichen Yang, Boyang Deng, Junjie Yan, Wei Wu, Jing Shao, and Cheng-Lin Liu. Blockqnn: Efficient block-wise neural network architecture generation. CoRR, abs/1808.05584, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  28. [28]

    Large-Scale Evolution of Image Classifiers

    Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V . Le, and Alex Kurakin. Large-scale evolution of image classifiers. CoRR, abs/1703.01041, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  29. [29]

    Hierarchical Representations for Efficient Architecture Search

    Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. CoRR, abs/1711.00436, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    DARTS: Differentiable Architecture Search

    Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: differentiable architecture search. CoRR, abs/1806.09055, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  31. [31]

    Neural architecture optimization

    Renqian Luo, Fei Tian, Tao Qin, and Tie-Yan Liu. Neural architecture optimization. CoRR, abs/1808.07233, 2018. 11 A PREPRINT - J UNE 27, 2019

    work page arXiv 2018

  32. [32]

    ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware

    Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. CoRR, abs/1812.00332, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  33. [33]

    Progressive Neural Architecture Search

    Chenxi Liu, Barret Zoph, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan L. Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. CoRR, abs/1712.00559, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  34. [34]

    Williams

    Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. In Machine Learning, pages 229–256, 1992

    work page 1992

  35. [35]

    Neural Combinatorial Optimization with Reinforcement Learning

    Irwan Bello, Hieu Pham, Quoc V . Le, Mohammad Norouzi, and Samy Bengio. Neural combinatorial optimization with reinforcement learning. CoRR, abs/1611.09940, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  36. [36]

    Geometric GAN

    Jae Hyun Lim and Jong Chul Ye. Geomtric gan. arXiv preprint arXiv:1705.02894, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  37. [37]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  38. [38]

    Learning to draw samples with amortized stein variational gradient descent

    Yihao Feng, Dilin Wang, and Qiang Liu. Learning to draw samples with amortized stein variational gradient descent. In Proceedings of the Thirty-Third Conference on Uncertainty in Artificial Intelligence, UAI 2017, Sydney, Australia, August 11-15, 2017, 2017

    work page 2017

  39. [39]

    Stacked Generative Adversarial Networks

    Xun Huang, Yixuan Li, Omid Poursaeed, John E. Hopcroft, and Serge J. Belongie. Stacked generative adversarial networks. CoRR, abs/1612.04357, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  40. [40]

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a nash equilibrium. CoRR, abs/1706.08500, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  41. [41]

    Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields

    Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, Martin Heusel, Hubert Ramsauer, and Sepp Hochreiter. Coulomb gans: Provably optimal nash equilibria via potential fields. CoRR, abs/1708.08819, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  42. [42]

    Improving generative adversarial networks with denoising feature matching

    David Warde-Farley and Yoshua Bengio. Improving generative adversarial networks with denoising feature matching. In ICLR, 2017

    work page 2017

  43. [43]

    Class-Splitting Generative Adversarial Networks

    Guillermo L. Grinblat, Lucas C. Uzal, and Pablo M. Granitto. Class-splitting generative adversarial networks. CoRR, abs/1709.07359, 2017. 12

    work page internal anchor Pith review Pith/arXiv arXiv 2017