# Population-Based Multi-Objective Training of Discriminators for Semi-Supervised GANs

arXiv: https://arxiv.org/abs/2607.01907 · HTML: https://mlsanity.com/paper/2607.01907/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 9 of its 23 references that are in the map.
- First posted on arXiv 2 Jul 2026.

## Builds on (most cited first, 9 of 9)

- [Generative Adversarial Networks](https://mlsanity.com/paper/2203.00667/) (2022)
- [Good Semi-supervised Learning that Requires a Bad GAN](https://mlsanity.com/paper/1705.09783/) (2017)
- [Evolutionary Generative Adversarial Networks](https://mlsanity.com/paper/1803.00657/) (2018)
- [COEGAN: Evaluating the Coevolution Effect in Generative Adversarial Networks](https://mlsanity.com/paper/1912.06180/) (2019)
- [Spatial Evolutionary Generative Adversarial Networks](https://mlsanity.com/paper/1905.12702/) (2019)
- [Data Dieting in GAN Training](https://mlsanity.com/paper/2004.04642/) (2020)
- [Signal Propagation in a Gradient-Based and Evolutionary Learning System](https://mlsanity.com/paper/2102.08929/) (2021)
- [Generate more than one child in your co-evolutionary semi-supervised learning GAN](https://mlsanity.com/paper/2504.20560/) (2025)
- [Multi-population GAN Training: Analyzing Co-Evolutionary Algorithms](https://mlsanity.com/paper/2507.13157/) (2025)

## Abstract (arXiv)

Semi-supervised generative adversarial networks (SSL-GANs) can exploit large unlabeled datasets while retaining a classifier in the discriminator, but their training is often unstable. This paper proposes a population-based evolutionary training strategy in which discriminator learning is formulated as a multi-objective optimization problem. Instead of aggregating the supervised and unsupervised components of the SSL objective into a single scalar loss, the method maintains a population of discriminators ranked by Pareto dominance, enabling the exploration of different trade-offs between classification accuracy and real/fake discrimination. This formulation aims to improve both roles of SSL-GANs: learning accurate classifiers and training generators capable of producing realistic samples. We analyze several variants, including an elitist strategy and a mono-objective ablation, to assess the role of multi-objective selection. Experiments on MNIST with limited labels show improved training robustness compared to SSL-GAN and CE-SSL-GAN state-of-the-art baselines, while the elitist variant consistently achieves the highest classification accuracy.