Practical Synthesis of Mixed-Tailed Data with Normalizing Flows
Abstract: Capturing the correct tail behavior is difficult, yet essential for a faithful generative model. In this work, we provide an improved framework for training flows-based models with robust capabilities to capture the tail behavior of mixed-tail data. We propose a combination of a tail-flexible base distribution and a robust training algorithm to enable the flow to model heterogeneous tail behavior in the target distribution. We support our claim with extensive experiments on synthetic and real world data.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: - Algorithm in appendix B fixed
- Added paragraph at the end of appendix b explaining that selective gen has not been used in the main paper.
- Moved references to use of GM in literature to the beginning of 4.2.1.
- Notation section fixed
- Removed mention of platykurtic and leptokurtic-ness of distributions in 4.1
- Figure 1 moved to the top of page 2
- Fixed table 2 arrow direction
- Mention of the detailed description of the metrics in the appendix added to section 5
- Figure 2 moved to Section 4.2.1 and modified (caption and explanation in the paper) for better clarity
- Citations fixed
For camera ready version:
- Added future work suggestion related to selective generation.
- Added an overall introduction of the selective generation process to Appendix B.
- Added acknowledgement section
- Author information, date, review link added for the camera-ready version
Assigned Action Editor: ~Alain_Durmus1
Submission Number: 2988
Loading