New method accurately estimates MMD variance for TimeGAN training

Researchers have developed a new method for accurately and efficiently estimating the variance of the Maximum Mean Discrepancy (MMD), a challenge particularly with unbalanced sample sizes. The proposed approach provides a finite-sample unbiased estimator and, by employing a recursive prefix-suffix accumulation scheme for the Laplace kernel, reduces computational complexity to O(N log N) with O(N) memory. This method has been validated for its theoretical exactness and numerical stability, proving effective in monitoring distributional convergence during the training of Time-series Generative Adversarial Networks (TimeGAN). AI

IMPACT Improves training stability for time-series generative models by providing a more accurate variance estimator.

RANK_REASON This is a research paper detailing a new statistical method with applications in generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• CORE Recommender
• Hugging Face
• Laplace kernel
• Maximum Mean Discrepancy
• Shijie Zhong
• TimeGAN
• Time-series Generative Adversarial Networks

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →