Evolutionary architecture search for generative adversarial networks using an aging mechanism-based strategy.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) are pivotal in AI for image generation.
• Traditional GANs suffer from training stability issues due to manual architecture design.

Purpose of the Study:

• To develop an automated method for designing stable and high-performing GAN architectures.
• To address the training stability challenges inherent in traditional GAN models.

Main Methods:

• Introduced Evolutionary Neural Architecture Search (ENAS) for GANs, named EAMGAN.
• Employed a one-shot model automating GAN architecture design.
• Integrated Operation Importance Metric (OIM) and an aging mechanism for stability and optimized search.
• Utilized a non-dominated sorting algorithm for Pareto-optimal solutions and diversity.

Main Results:

• EAMGAN demonstrated competitive efficiency and performance on benchmark datasets.
• Achieved Inception Score (IS) of 8.83±0.13 and Fréchet Inception Distance (FID) of 9.55 on CIFAR-10 using only 0.66 GPU days.
• Showcased robust portability across STL-10, CIFAR-100, and ImageNet32 datasets.

Conclusions:

• EAMGAN effectively automates GAN architecture search, significantly improving training stability.
• The proposed method offers a highly efficient and performant solution for advanced image generation.
• EAMGAN's architecture search demonstrates broad applicability and robust performance across diverse datasets.