Towards Healing the Blindness of Score Matching

TL;DR

This paper introduces a new family of divergences designed to address the blindness problem in score-based methods, improving density estimation for multi-modal distributions in machine learning.

Contribution

The authors propose a novel divergence family that mitigates blindness in score matching, enhancing performance in density estimation tasks.

Findings

Improved density estimation accuracy for multi-modal distributions

Demonstrated effectiveness of the new divergence over traditional methods

Addressed the blindness problem in score-based divergences

Abstract

Score-based divergences have been widely used in machine learning and statistics applications. Despite their empirical success, a blindness problem has been observed when using these for multi-modal distributions. In this work, we discuss the blindness problem and propose a new family of divergences that can mitigate the blindness problem. We illustrate our proposed divergence in the context of density estimation and report improved performance compared to traditional approaches.