Finding maxima of Gaussian Sum-Product Networks

This thesis is about finding maxima of Sum-Product Networks (SPNs). SPNs are expressive statistical deep models that efficiently represent complex probability distributions. They encode context-specific independence among random variables and enable exact marginal and conditional probability inferen...

Descripción completa

Detalles Bibliográficos
Autor: Madeira, Tiago
Tipo de recurso: tesis de maestría
Estado:Versión publicada
Fecha de publicación:2023
País:Brasil
Institución:Universidade de São Paulo (USP)
Repositorio:Biblioteca Digital de Teses e Dissertações da USP
Idioma:inglés
OAI Identifier:oai:teses.usp.br:tde-18092023-103415
Acceso en línea:https://www.teses.usp.br/teses/disponiveis/45/45134/tde-18092023-103415/
Access Level:acceso abierto
Palabra clave:Aprendizagem de máquina
Busca de modas
Gaussian mixture models
Machine learning
Mode finding
Modelos de misturas Gaussianas
Modelos probabilísticos
Probabilistic models
Redes Soma-Produto
Sum-Product Networks
Descripción
Sumario:This thesis is about finding maxima of Sum-Product Networks (SPNs). SPNs are expressive statistical deep models that efficiently represent complex probability distributions. They encode context-specific independence among random variables and enable exact marginal and conditional probability inference in linear time. The research explores Gaussian SPNs (GSPNs), which are continuous SPNs with Gaussian distributions at their leaves. GSPNs provide compact representations of Gaussian Mixture Models (GMMs) with many components. The relationship between GSPNs and GMMs has been largely unexplored in the literature, particularly regarding mode-finding techniques. The problem of finding modes in Gaussian mixtures is challenging, and existing techniques involve hill-climbing algorithms. However, there is limited research discussing modes in the context of SPNs. The objective of this work is to investigate and establish a framework for identifying modes in GSPNs. This is accomplished by developing an algorithm that employs an EM-style fixed-point iteration method for mode finding in GSPNs. The algorithm is presented in detail, accompanied by a formal proof of its correctness. Two applications for it are discussed: Maximum-A-Posteriori inference and modal clustering. Some experimental results are provided to evaluate the effectiveness of the proposed approach.