Conditional Sampling and Density Estimation with Triangular Convex Flows Público

Wang, Zheyu (Spring 2023)

Permanent URL: https://etd.library.emory.edu/concern/etds/gq67js59q?locale=es
Published

Abstract

We introduce Triangular Convex Flows (TC-Flow), a method for learning conditional probability distributions given samples from the joint probability distributions. Unlike previous methods that rely on constructing monotone triangular transport maps through soft penalties and partial integration, TC-Flow uses a novel map parameterization based on the input gradient of scalar-valued fully and partially input convex neural networks (FICNN and PICNN). The approach guarantees monotone maps without requiring specific network weights and ensures optimal maps under optimal transport theory by approximating Kantorovich potentials. During training, we parallelize the process over map components and minimize the expected negative log-likelihood of the samples. We demonstrate the effectiveness of TC-Flow on synthetic two-dimensional datasets and compare it to existing triangular maps model on high-dimensional benchmark problems. Our numerical experiments show that TC-Flow is competitive with the state-of-the-art model in both expressiveness and scalability.

Table of Contents

1 Introduction ........................ 1

1.1 Contributions and Outline ........................ 3

2 Theoretical and Modeling Background ........................ 4

2.1 Optimal Transport Maps......................... 4

2.2 Monotone Triangular Transport Maps.................. 5

2.3 Maximum Likelihood Estimation .................... 6

2.4 Adaptive Transport Map Model..................... 8

3 Formulation and Analysis of Triangular Convex Flows ........................ 10

3.1 Input Convex Neural Networks ..................... 10

3.2 ICNN Triangular Maps.......................... 11

3.3 ICNN Block Triangular Maps ...................... 12

3.4 Training Problem............................. 14

3.5 Inversion Problem............................. 15

4 Numerical Experiments ........................ 17

4.1 Two Dimensional Synthetic Data .................... 17

4.2 Tabular Dataset.............................. 19

5 Discussions and Future Directions ........................ 22

6 Concluding Remarks ........................ 24

Bibliography ........................ 25 

About this Honors Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
School
Department
Degree
Submission
Language
  • English
Research Field
Palabra Clave
Committee Chair / Thesis Advisor
Committee Members
Última modificación

Primary PDF

Supplemental Files