Research map / Image generation / GANs
Neuroevolution of a Multi-Generator GAN (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a neuroevolution method that jointly trains a population of generators against a shared discriminator, combining backpropagation (adversarial) and evolutionary hypermutation steps, and evaluates variants on MNIST against DCGAN.
Problem
GANs suffer from mode collapse, vanishing gradients, and limited control over generated samples; prior EA-GAN work uses evolution and backpropagation in isolation rather than jointly tuning the same GAN parameters.
Approach
A multi-generator GAN where each generation runs adversarial backpropagation on all generators and the discriminator, then an evolutionary step that hypermutates (Gaussian noise on intermediate conv layers) the worse half of generators and keeps the best one, so that backpropagation exploits and evolution explores.
Key results
- Variants of the proposed method, except 1G2CGAN, generate images comparable in quality to DCGAN on MNIST while using 5 epochs per generation over 5 generations versus 25 epochs for DCGAN.
- Mutating only the last convolution layer yields better results than mutating multiple layers, and it stabilises GAN convergence.
- Using multiple generators in both adversarial and evolutionary optimisation yields better-quality images but tends toward repetitive content, possibly mode collapse.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the 1G2CGAN strategy of mutating the last two convolutional layers (ablation) | Proposed multi-generator GAN · MNIST | image quality | Hurt |
| Mutate one layer rather than multiple layers (ablation) | Proposed multi-generator GAN · MNIST | image quality | Helped |
| Use multiple generators in both adversarial and evolutionary optimization (ablation) | Proposed multi-generator GAN · MNIST | image quality | Helped |
| Use multiple generators in both adversarial and evolutionary optimization (ablation) | Proposed multi-generator GAN · MNIST | repetitive content | Hurt |
| Mutate only one layer (ablation) | Proposed multi-generator GAN · MNIST | GAN convergence stability | Helped |
Surprising findings
- Using multiple generators in both optimisation routines produces repetitive outputs, possibly indicating mode collapse.
Limitations
- Evaluation restricted to MNIST, a simple dataset, as a preliminary study.
- Only qualitative visual comparison of generated images; no quantitative metrics such as FID reported in the main text.
- Merging adversarial and evolutionary optimisation is described as complex; mode collapse observed with multiple generators.
Experimental setup
Datasets: MNIST
Benchmarks: MNIST
Baselines: DCGAN
Builds on: GAN DCGAN
Closest papers by meaning
- Are GANs Created Equal? A Large-Scale Study (NeurIPS 2018 · similarity 0.86)
- OPT-GAN: A Broad-Spectrum Global Optimizer for Black-Box Problems by Learning Distribution (AAAI 2023 · similarity 0.86)
- The Numerics of GANs (NeurIPS 2017 · similarity 0.86)
- Variational Annealing of GANs: A Langevin Perspective (ICML 2019 · similarity 0.86)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.85)
- RG-GAN: Dynamic Regenerative Pruning for Data-Efficient Generative Adversarial Networks (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.