Scholarlink

Multiple adversary networks represent a recurring extension in generative adversarial learning, leveraging discriminator ensembles for improved model performance and training stability. However, determining when discriminator ensembles provide substantial benefits remains unclear. We systematically apply this approach across image, tabular, and spectral data, demonstrating significant improvements with minimal implementation complexity - notably reducing Frechet Inception Distance scores by up to 40.6% for image generation. Our comprehensive study, spanning 11 distinct use-cases, pioneers the underexplored realm of multi-adversarial techniques for tabular and spectral data synthesis. We identify gradient orthogonality within discriminator ensembles as the primary driver of performance gains. Our findings provide practical guidance on when to implement multi-adversarial approaches, complemented by gradient-based measures for monitoring ensemble dynamics and quantifiable performance expectations across various architectures.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

When to Use Discriminator Ensembles? Cross-Domain Evidence of Multi-Adversarial Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Benjamin Schindler,
  • Melle Mendikowski,
  • Thomas Schmid,
  • Sam Verboven

摘要

Multiple adversary networks represent a recurring extension in generative adversarial learning, leveraging discriminator ensembles for improved model performance and training stability. However, determining when discriminator ensembles provide substantial benefits remains unclear. We systematically apply this approach across image, tabular, and spectral data, demonstrating significant improvements with minimal implementation complexity - notably reducing Frechet Inception Distance scores by up to 40.6% for image generation. Our comprehensive study, spanning 11 distinct use-cases, pioneers the underexplored realm of multi-adversarial techniques for tabular and spectral data synthesis. We identify gradient orthogonality within discriminator ensembles as the primary driver of performance gains. Our findings provide practical guidance on when to implement multi-adversarial approaches, complemented by gradient-based measures for monitoring ensemble dynamics and quantifiable performance expectations across various architectures.