Source-linked AI summary
Evolutionary Generative Adversarial Networks
Chaoyue Wang, Chang Xu, Xin Yao, Dacheng Tao
TL;DR
GANs provide effective generative modeling but often face unstable training and mode collapse. E-GAN evolves a population of generators using multiple adversarial objectives as mutations, evaluates sample quality and diversity, and preserves strong offspring. Experiments across several image-generation tasks report improved stability and convincing performance.
Problem
GAN training often suffers from instability, mode collapse, and sensitivity to hyper-parameters or architectures.
Method
E-GAN evolves a population of generators under a discriminator environment, using different adversarial objectives as mutations and selecting offspring by quality and diversity.
Results
Experiments on several image-generation tasks show improved training stability and convincing generative performance.
Takeaways & Limitations
E-GAN preserves the best offspring from different objectives, addressing limitations of individual adversarial training objectives.
Takeaways & Limitations
E-GAN takes more time for each iteration than conventional alternatives.
Abstract
from arXiv · showhide
Generative adversarial networks (GAN) have been effective for learning generative models for real-world data. However, existing GANs (GAN and its variants) tend to suffer from training problems such as instability and mode collapse. In this paper, we propose a novel GAN framework called evolutionary generative adversarial networks (E-GAN) for stable GAN training and improved generative performance. Unlike existing GANs, which employ a pre-defined adversarial objective function alternately training a generator and a discriminator, we utilize different adversarial training objectives as mutation operations and evolve a population of generators to adapt to the environment (i.e., the discriminator). We also utilize an evaluation mechanism to measure the quality and diversity of generated samples, such that only well-performing generator(s) are preserved and used for further training. In this way, E-GAN overcomes the limitations of an individual adversarial training objective and always preserves the best offspring, contributing to progress in and the success of GANs. Experiments on several datasets demonstrate that E-GAN achieves convincing generative performance and reduces the training problems inherent in existing GANs.
1 Introduction
GANs learn generative models through adversarial training, but training can suffer from instability, mode collapse, and sensitivity to configurations. E-GAN addresses these difficulties by evolving generators under multiple adversarial objectives and preserving strong offspring.
- GANs learn generative models by training a generator to synthesize realistic samples and a discriminator to distinguish real from generated data.
- GAN training can produce vanishing or unreliable gradients, mode collapse, and poor results under unsuitable hyper-parameters or architectures.
- Recent GAN research develops alternative adversarial objectives that measure distributional differences under different metrics.
- E-GAN treats the discriminator as an environment and evolves a population of generators using different adversarial objectives as mutations.
- E-GAN combines different objectives and preserves well-performing offspring, improving training stability and generative performance across several datasets.
2 Related Works
Related work covers GAN formulations and efforts to improve their stability, alongside evolutionary algorithms for population-based variation and selection. Figure 1 contrasts conventional two-player GAN training with E-GAN’s evolutionary framework.
- GANs: GANs alternate updates to a generator and discriminator and have produced strong samples in many generative tasks.
- GAN improvements: Existing GAN variants address training problems through heuristic objectives, architectural changes, and Wasserstein-based objectives.
- E-GAN framework: Figure 1 contrasts a single generator–discriminator adversarial game with a population of generators evolving through variation, evaluation, and selection.
- Evolutionary algorithms: Evolutionary algorithms represent candidate solutions as individuals, generate offspring through variation, and select solutions according to fitness.
- Evolutionary deep learning: Evolutionary methods have been applied to deep learning for hyper-parameter optimization, architecture design, and neural-network optimization.
3 Method
The method section introduces E-GAN, its mutations and evaluation mechanism, and the complete evolutionary training process.
- The method section presents E-GAN by reviewing GANs, introducing the algorithm, explaining its mutations and evaluation mechanism, and summarizing training.
3.1 Generative Adversarial Networks
GANs formulate generation as a two-player minimax game in which a generator maps noise to samples and a discriminator distinguishes real from generated data. Existing GANs implement this procedure with different adversarial objectives.
- A generator maps noise z to generated data G(z), while a discriminator distinguishes real samples from generated samples.
- The adversarial procedure alternates generator and discriminator training under different adversarial objective functions.
3.2 Evolutionary Algorithm
E-GAN treats generators as an evolving population adapting to a discriminator environment through variation, fitness-based evaluation, selection, and discriminator updates.
- Each generator acts as a possible solution, while the discriminator provides the environment to which the population adapts.
- Variation: Variation creates multiple children from each parent by applying different mutations to copied generator individuals.
- Evaluation: A fitness function evaluates each child’s performance using the current discriminator environment.
- Selection: Selection removes the lowest-fitness children, leaving the remaining individuals available as parents for the next iteration.
- After each evolutionary step, the discriminator is updated to distinguish real samples from samples generated by the evolved generators.
- The updated discriminator continually supplies adaptive losses that drive the generator population toward better solutions.
3.3 Mutations
E-GAN uses several generator objectives as mutation operators, combining their distinct gradient and distribution-matching behaviors within an evolutionary framework.
- Asexual reproduction produces children through different mutation operators corresponding to training objectives that narrow distribution distances from different perspectives.
- Minimax mutation: The minimax mutation minimizes Jensen-Shannon divergence but can suffer vanishing gradients when distribution supports lie on separate manifolds.
- Minimax mutation: When the discriminator cannot fully distinguish real from fake samples, the minimax mutation provides effective gradients that narrow the distribution gap.
- Heuristic mutation: The heuristic mutation maximizes the discriminator’s mistaken-classification log probability and avoids saturation when generated samples are rejected.
- Heuristic mutation: The heuristic mutation can nevertheless cause instability because its optimal-discriminator form includes inverted KL minus two JSDs.
- Least-squares mutation: The least-squares mutation is non-saturating when the discriminator recognizes generated samples, helping avoid vanishing gradients.
- One discriminator is sufficient to provide adaptive losses for the described mutations because LSGAN’s optimal discriminator is equivalent to theirs.
3.4 Evaluation
E-GAN evaluates generator children using fitness measures for sample quality and diversity, balancing these criteria to guide selection and improve training behavior.
- The evaluation function measures individual quality to determine evolutionary direction and select generator children.
- Quality: Quality fitness is estimated from the discriminator’s average output on images produced by a generator.
- Quality: The discriminator is continually upgraded, allowing its assessment to reflect generator quality at each evolutionary step.
- Diversity: Diversity evaluation targets mode collapse by using discriminator-gradient behavior to assess optimization stability and generated-sample diversity.
- Diversity: Small discriminator gradients correspond to higher diversity scores, with samples spreading sufficiently to suppress mode collapse and smooth discriminator changes.
- The overall fitness combines quality and diversity through a nonnegative γ that balances the two measurements.
3.5 E-GAN
E-GAN evolves generator populations with multiple objectives and preserves high-performing offspring, integrating complementary objectives while suppressing their individual training limitations.
- Generators form an evolutionary population, while the discriminator acts as an environment that guides different objective-based mutations.
- Only well-performing children survive and participate in future adversarial training according to survival of the fittest.
- E-GAN integrates advantages of different adversarial objectives while suppressing limitations such as vanishing gradients and mode collapse.
4 Experiments
Experiments across synthetic and image datasets show that E-GAN improves mode coverage, training stability, convergence, and image quality across varied architectures.
- Synthetic Datasets and Mode Collapse: After 50K iterations, E-GAN more accurately fits 8- and 25-Gaussian target distributions than individual adversarial objectives, which exhibit mode collapse.The evaluation mechanism selects well-performing offspring and promotes the population toward better evolutionary directions.
- CIFAR-10 and Inception Score: On CIFAR-10, E-GAN reaches higher inception scores with fewer training steps and maintains comparable stability at convergence.The comparison uses the same DCGAN-based architecture across methods.
- CIFAR-10 and Inception Score: E-GAN outperforms other GANs while achieving comparable wall-clock convergence speed despite taking more time per iteration.The comparison includes WGAN and WGAN-GP.
- CIFAR-10 and Inception Score: E-GAN adapts mutation selection during training: heuristic and least-square mutations dominate early, while minimax mutations become more frequent after 20K steps.This selection pattern is associated with avoiding ineffective gradients early and improving stability near convergence.
- LSUN and Architecture Robustness: Across challenging LSUN architectures, E-GAN generates reasonable results even when other tested methods fail.The tested methods are DCGAN, LSGAN, WGAN, WGAN-GP, and E-GAN, trained for 200K iterations.
- CelebA and Space Continuity: On CelebA, interpolated generated faces change seamlessly across meaningful attributes, indicating space continuity and no observed mode collapse.The evaluated attributes include gender, expression, hairstyle, and age.
5 Conclusion
The paper presents E-GAN, an evolutionary framework that adapts a population of generators to a discriminator while preserving the best offspring. Experiments show improved training stability and convincing performance across image-generation tasks.
- Conclusion: E-GAN evolves a population of generators in response to the discriminator and preserves the best offspring after each iteration.The framework is designed to reduce training difficulties and improve generative performance.
- Conclusion: Experiments show that E-GAN improves GAN training stability and achieves convincing performance across several image-generation tasks.The paper identifies further exploration of discriminator–population relationships and generative performance as future work.