acceptodds
Under review as a conference paper at ICLR 2027

Epistemic Generative Adversarial Networks

Abstract

Generative models, particularly Generative Adversarial Networks (GANs), often suffer from a lack of output diversity, frequently generating similar samples rather than a wide range of variations. This paper introduces a novel generalization of the GAN loss function based on Dempster-Shafer theory of evidence, applied to both the generator and discriminator. Additionally, we propose an architectural enhancement to the generator that enables it to predict a mass function for each image pixel. This modification allows the model to quantify uncertainty in its outputs and leverage this uncertainty to produce more diverse and representative generations. Experimental evidence shows that our approach not only improves generation variability but also provides a principled framework for modeling and interpreting uncertainty in generative processes.

Primary area:
generative models
Keywords:
Generative Adversarial Networks, Uncertainty Quantification, Computer Vision, Deep Generative Models, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NiLvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Epistemic GAN, a framework that incorporates Dempster–Shafer evidence theory into generative adversarial networks to improve diversity in generation. The discriminator is modified to predict separate belief values for real and fake samples, allowing the remaining belief mass to represent uncertainty or ignorance. The generator is also redesigned into two stages, where Dirichlet distributions are used to sample interval representations before producing the final images. The authors introduce corresponding loss terms to balance diversity and precision. Experiments on CelebA, CIFAR-10, and Food-101 show improvements in both FID and Vendi Score over a standard DCGAN baseline.

Critical strengths
  1. The idea of introducing evidential uncertainty into GAN training is interesting. Rather than addressing mode collapse solely through architectural changes or regularization, the paper explores whether explicitly representing uncertainty could help improve generation diversity. I find this perspective worth investigating, particularly the attempt to incorporate belief functions into both the generator and discriminator.
  2. The empirical results are encouraging, although still preliminary. The method consistently improves FID and Vendi Score over the standard GAN baseline across three datasets. I also appreciate the additional ablation studies and computational overhead analysis, which help provide insight into the different components and their practical costs.
Critical weaknesses
  1. I am not fully convinced that the proposed framework actually captures meaningful epistemic uncertainty, or that the improvements can be attributed to the evidential formulation.
    My main concern is the gap between the theoretical motivation and what the proposed objectives actually encourage. For the discriminator, the additional belief outputs allow an "ignorance" gap in principle. However, considering the discriminator loss as written, its pointwise optimum for a fixed generator appears to satisfy breal+bfake=1b_{\mathrm{real}}+b_{\mathrm{fake}}=1, effectively eliminating this gap. The additional parameterization may still affect training dynamics, but the paper does not provide evidence showing when meaningful ignorance emerges or how it benefits optimization.
    I have a similar concern about the generator. The Dirichlet-based interval sampling introduces additional stochasticity, but it remains unclear why this should be interpreted as epistemic uncertainty rather than simply another form of noise injection. In particular, the paper does not demonstrate that the predicted uncertainty reflects meaningful properties of the generated samples or that it is responsible for improved mode coverage.
    There also seems to be an inconsistency in the generator loss: the Dirichlet variance term has a positive coefficient, even though the text states that the objective encourages higher variance. Under the usual loss minimization convention, this term would have the opposite effect. I think this needs clarification, as it directly concerns the proposed diversity mechanism.
    Overall, these issues make it difficult to assess whether the evidential interpretation is an essential contribution of the method or mainly an alternative interpretation of a stochastic GAN architecture.

  2. The experimental evidence is not yet sufficient to support the broader claims about improving generation diversity and mitigating mode collapse.
    The experiments rely almost entirely on comparisons with DCGAN, which is a rather old and relatively weak baseline. I understand the motivation for keeping the backbone fixed to isolate the proposed changes, but this alone does not establish whether the approach offers meaningful advantages over existing techniques for improving GAN diversity. The models are also trained for only five epochs, making it difficult to determine whether the observed improvements persist after convergence.
    The ablation studies are useful, but they do not completely resolve this concern. For example, the full model achieves neither the best FID nor the highest Vendi Score among the architectural variants in Table 5. More importantly, the paper does not sufficiently disentangle the benefits of evidential modeling from those of additional stochasticity or changes in generator capacity.
    I am also hesitant to interpret higher Vendi Scores as direct evidence of reduced mode collapse. Increased sample diversity does not necessarily imply better coverage of the target distribution. More direct evaluations of mode coverage, together with stronger baselines and statistical variation across runs, would make the conclusions substantially more convincing.

Questions for authors
  1. Could the authors clarify the relationship between the proposed discriminator objective and the standard GAN objective? In particular, the pointwise optimal discriminator appears to satisfy breal+bfake=1b_{\mathrm{real}}+b_{\mathrm{fake}}=1, leaving no ignorance mass. Is this interpretation correct? If so, is the intended benefit primarily related to transient optimization dynamics? It would be helpful to see empirical evidence, such as how the ignorance mass evolves during training and whether it leads to different or more useful gradient signals.

  2. How can we distinguish the benefits of epistemic uncertainty modeling from those of simply introducing additional stochasticity? For example, would a generator with a conventional stochastic bottleneck or noise injection, but without the proposed belief-theoretic interpretation, achieve similar improvements? Also, could the authors clarify the sign of the Dirichlet variance term in the generator loss? As written, minimizing the objective seems to discourage rather than encourage variance.

  3. Could the authors provide more direct evidence that the method improves mode coverage? While the Vendi Score is a useful measure of diversity, it does not necessarily indicate whether previously missing modes are recovered. An explicit mode coverage analysis, or an examination of how the predicted uncertainty relates to sample diversity, would help establish the claimed mechanism.

  4. How robust are the improvements to longer training and stronger baselines? Since all models are trained for only five epochs, I wonder whether the performance differences would remain after convergence. Comparisons with other diversity-oriented GAN objectives, preferably under matched model capacity, would also help establish the practical value of the proposed framework. Additionally, reporting the mean and standard deviation of FID and Vendi Score across multiple runs would make the relatively modest improvements easier to assess.

Reviewer_v8VhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on improving the diversity of image generation in generative adversarial networks. The authors find that GANs often suffer from mode collapse and lack a principled framework for modeling uncertainty in the generative process. They therefore propose Epistemic GAN, which replaces the discriminator output with belief functions over real and fake samples and lets the generator predict region-wise Dirichlet mass functions. To validate their method, they conduct several experiments, including comparisons with a standard DCGAN on CelebA, CIFAR-10, and Food-101, hyperparameter and component ablations, and a training-cost analysis. The experimental results demonstrate that Epistemic GAN achieves lower FID and higher Vendi scores than the standard DCGAN.

Critical strengths
  1. This paper proposes Epistemic GAN, which introduces Dempster-Shafer belief functions into both the discriminator and generator of a GAN.

  2. The authors introduce region-wise Dirichlet mass prediction in the generator, which adds only about 1.5% training time per epoch.

  3. They provide ablations on the loss weights and on each evidential component.

Critical weaknesses
  1. My main concern is that the uncertainty claims are not validated. The Abstract and Section 4 state that the generator provides interpretable region-wise uncertainty maps that highlight object boundaries and texture transitions, and the first contribution claims improved explainability. However, the paper neither shows such maps nor evaluates any uncertainty estimate, and all experiments report only FID and Vendi scores. The epistemic component that gives the method its name therefore remains unsupported.

  2. The experiments do not establish how the method compares with current generative models. All results use a DCGAN backbone trained for five epochs, with every dataset resized to 64 × 64, including the 32 × 32 CIFAR-10 images. No modern GAN backbone is tested, although Section 5.3 presents the method as a drop-in replacement for standard architectures, and Section 5.1 deliberately excludes the diversity-promoting methods reviewed in Section 2. These non-standard settings also prevent comparison with published results, so it remains unknown whether the gains hold at the quality level of current GANs.

  3. The gains are not shown to stem from evidential reasoning. The discriminator loss in Section 4.2 amounts to two binary cross-entropy heads with mirrored labels plus a soft penalty on breal+bfake≤1b_{real} + b_{fake} \le 1, while the ignorance mass plays no further role. In the generator, sampling a Dirichlet distribution per region injects randomness under fixed parameters, which is output stochasticity rather than epistemic uncertainty and may increase diversity regardless of its belief interpretation.

  4. Key parts of the method are under-specified. Section 4.1 does not explain how a sample from the three-parameter Dirichlet distribution becomes an interval, what the second generator module receives, or how gradients pass through the sampling process. In the generator loss, the Dirichlet variance is not a scalar and appears with a positive sign, so minimizing the loss would reduce it, contrary to the stated aim of maximizing variance.

  5. The evaluation protocol is incomplete. Section 5.1 does not state the number of samples used for FID or the feature space and kernel used to define the Vendi score, although Vendi values depend on these choices. This limits the interpretation of the diversity gains reported in Table 1.

Questions for authors
  1. Could the authors visualize the predicted uncertainty maps and provide evidence that they reflect meaningful properties of the generated images?

  2. Could the authors apply the evidential formulation to a modern GAN backbone, such as StyleGAN2 ("Analyzing and Improving the Image Quality of StyleGAN," CVPR 2020), under standard benchmark protocols, and compare it with representative diversity-promoting methods from Section 2?

  3. Could the authors add controls that inject matched per-region noise into a standard generator and replace the evidential discriminator with two unconstrained heads? This would help isolate the contribution of the Dempster-Shafer formulation.

  4. Could the authors specify the Dirichlet-to-interval mapping, the input to the second generator module, and the gradient estimator, and clarify the definition and sign of the variance term?

  5. Could the authors specify the number of generated samples used for each metric and the embedding and kernel used for the Vendi score?

Reviewer_jymcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Epistemic GAN, which leverages Dempster–Shafer evidence theory to improve the generative diversity of GANs and provides a representation of local uncertainty. The discriminator outputs the confidence levels for real and generated samples, respectively, interpreting the remaining quality as ignorance. The generator predicts the Dirichlet distribution in intermediate feature regions, constructs intervals through sampling, and then decodes them into images by subsequent networks. The training objective incorporates variance and interval width regularization terms in addition to the adversarial loss to adjust for both diversity and generation accuracy. Based on DCGAN, the authors report improvements in FID and Vendi Score on datasets such as CelebA, and provide ablation of loss weights, generator and discriminator components, and qualitative results for conditional generation of Cityscapes.

Critical strengths

Consistent improvements were observed in FID and Vendi Score across all three datasets, demonstrating effectiveness not limited to a single dataset. For example, the FID for Food-101 decreased from 33.76 to 29.1, and for Vendi, it increased from 12.78 to 13.82. Appendix tables also examine the loss weights and the cases where only the generator or discriminator was modified. While the complete model did not outperform all individual metrics, these results at least demonstrate the varying impacts of different components on quality and diversity, providing preliminary empirical support for the method. The method attempts to control generation randomness at the local feature level, rather than relying solely on latent variables at the input, a design that is quite attractive. Modifications to the discriminator were relatively simple, primarily involving the output layer and the loss. Table 2 reports an increase in training time per epoch on CelebA from 75.91 seconds to 77.10 seconds, indicating minimal additional computational overhead under the tested configuration. If the gains can be consistently maintained, this modification has practical value.

Critical weaknesses

The experiments are insufficient to support the core conclusion of "modeling cognitive uncertainty," nor do they fully determine the source of the diversity gains. Existing metrics primarily measure generation quality and inter-sample differences; Vendi itself does not depend on the true reference distribution, so an increase in score cannot alone indicate coverage of more real patterns. In Appendix Table 3, higher Vendi is accompanied by significantly worse FID, which the authors interpret as diversity resulting from noise or incoherent samples. Furthermore, the main experiment only compares with a DCGAN trained for five rounds, without reporting cross-seed fluctuations in FID/Vendi, and lacks a control on ordinary random layers on the same backbone. Therefore, it is currently difficult to determine whether the gains depend on the design of the evidence theory, and whether they still hold true after sufficient training. Section 4.1 does not provide a specific mapping from 3D Dirichlet samples to interval endpoints, nor does it explain how backpropagation is performed through sampling. Since the interval width itself participates in training, this information directly determines the meaning and reproducibility of the method.

Questions for authors

Regarding the explanation of uncertainty and payoff attribution, what direct evidence suggests that interval width or ignorance quality reflects a lack of knowledge, rather than just adjustable internal randomness? Can you provide a verification of uncertainty itself, comparing it with ordinary local noise injection under the same backbone, similar parameter count, and sufficient training conditions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.