Epistemic Generative Adversarial Networks
Muhammad Mubashar, Fabio Cuzzolin
TL;DR
This paper proposes a new GAN framework based on Dempster-Shafer theory that enhances output diversity and uncertainty quantification, leading to more varied and interpretable generated samples.
Contribution
It introduces a Dempster-Shafer based loss function and an architectural modification for the generator to predict pixel-wise uncertainty, improving diversity and interpretability.
Findings
Enhanced sample diversity in generated images
Quantified uncertainty improves interpretability
Framework provides a principled approach to uncertainty modeling
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.
Peer Reviews
Decision·Submitted to ICLR 2026
This paper introduces the concept of belief into adversarial training, which is interesting and merits further investigation.
1. My main concern lies in the disconnect between the proposed theoretical formulation and its practical implementation. Although the paper introduces the model using the language of belief and interval theory, these concepts are not meaningfully reflected in the training process. In the discriminator, the outputs $b_{real}$ and $b_{fake}$ are described as belief measures for real and fake samples, yet aside from a penalty enforcing $b_{real}+b_{fake} <=1$, they behave identically to a standard
* The underlying theory behind the proposed approach is interesting and could be of some interest to the community. * The explanation of the Dempster-Shafer theory is also fairly well done. * The modifications to the architecture and the loss function seem to be well motivated based on the above theory and its explanation * In experimental evaluation, using the same base architecture and compute constraints is laudable. * The datasets used for evaluation seem fine, more details on it in next sec
* While the underlying motivation and theoretical underpinnings are appropriate and understandable, I am not convinced that it's practical instantiation in the modifications to the discriminator and generator has been well justified or evaluated in the paper. * There are two distinct architectural changes proposed in the paper. I was expecting each of these changes to be evaluated separately, not only on the final metrics, but on samples and the loss themselves. * Specifically for the discrimina
1. Converting the discriminator output and the intermediate generator representation into a belief/mass function may be an original take on uncertainty-aware GAN training. The DS tutorial is clear and self-contained. 2. The two-stage generator with a Dirichlet proxy for interval hypotheses is well-motivated and concretely described, with helpful schematics. 3. The paper is generally well-written and easy to follow.
1. The paper compares mainly to a “standard” DCGAN (from 2015). It omits established diversity-oriented GAN variants and modern strong baselines (StyleGAN, or diffusion-based contenders using diversity metrics). Without these, it’s hard to isolate how much of the gain stems from the evidential machinery versus general architectural changes. 2. The paper lacks ablations for (i) evidential loss in the discriminator only, (ii) mass-predicting generator only, (iii) the Dirichlet interval mapping vs
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Explainable Artificial Intelligence (XAI) · Adversarial Robustness in Machine Learning