📄 中文摘要
AI在学习如何识别事物或进行排名时，通常使用一种称为沃瑟斯坦的评分方法。然而，在训练过程中，这种方法有时会发出偏差信号，导致模型朝错误的方向更新，从而影响生成图像或排名的质量。研究人员寻找解决方案，发现克拉梅距离是一种更简单的比较模型预测与实际结果的方法。克拉梅距离保留了旧评分的优点，同时避免了误导性的信号。研究表明，将这一理念应用于生成对抗网络（GAN）中，训练过程明显更平滑，生成的结果减少了奇怪的伪影，效果更佳。
克拉梅距离作为解决偏差沃瑟斯坦梯度的方案
出处: The Cramer Distance as a Solution to Biased Wasserstein Gradients
发布: 2026年2月17日
📄 English Summary
The Cramer Distance as a Solution to Biased Wasserstein Gradients
AI models often rely on a scoring method known as Wasserstein when learning to recognize objects or rank items. However, this method can sometimes produce biased signals during training, leading models to update incorrectly, which affects the quality of generated images or rankings. Researchers sought a solution and discovered that the Cramér distance offers a simpler way to compare model predictions with actual outcomes. This approach retains the advantages of previous scoring methods while avoiding misleading nudges. Implementing this concept within a type of AI known as Generative Adversarial Networks (GANs) resulted in noticeably smoother training, fewer artifacts, and overall better performance.