Manifold Matching via Deep Metric Learning for Generative Modeling

By M. Dai et al

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Methodology
3.1. Proposed Framework
3.2. Manifold Matching
3.3. Metric Learning
3.4. Objective Functions
4. Conclusion

Summary

The paper proposes a manifold matching approach for generative modeling, focusing on deep generative models and deep metric learning. It introduces a framework containing a distribution generator and a metric generator to match geometric descriptors between real and generated data sets using learned distances. The approach aims to optimize the measure and metric around a real data manifold and uses Triplet metric learning for proper metric selection. The paper discusses the implementation pipeline, training procedures, and objectives for unconditional image generation tasks and super-resolution tasks, demonstrating the feasibility and effectiveness of the proposed framework.