Smoothing the Score Function to Enhance Generalization in Diffusion Models
The episode discusses 'Smoothing the Score Function to Enhance Generalization in Diffusion Models,' a paper by Zhou, Zhang, and Wright. The hosts explain that generative models often suffer from memorization because their mathematical weights are too concentrated. They discuss two solutions—Noise Unconditioning and Temperature Smoothing—to improve the model's ability to create original content.