A benchmark for Maximum-a-Posteriori Inference algorithms in discrete Sum-Product Networks
The solution to Maximum-a-Posteriori Inference problems in Sum-Product Networks provides the most probable configuration of the Random Variables encoded in its structure; a key step in Probabilistic reasoning that can be used for many applications, such as image auto-completion. It has been proven t...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2021 |
| 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-19062021-063556 |
| Acceso en línea: | https://www.teses.usp.br/teses/disponiveis/45/45134/tde-19062021-063556/ |
| Access Level: | acceso abierto |
| Palabra clave: | Maximum-a-posteriori Modelos probabilísticos Probabilistic models Redes soma-produto Sum-product networks |
| Sumario: | The solution to Maximum-a-Posteriori Inference problems in Sum-Product Networks provides the most probable configuration of the Random Variables encoded in its structure; a key step in Probabilistic reasoning that can be used for many applications, such as image auto-completion. It has been proven that this problem is NP-Hard (even to approximate) in Sum-Product Networks. Multiple algorithms have been developed to reach either approximate or exact solutions to this problem, but the experiments have been limited. In this Dissertation, we provide descriptions, analysis, and a benchmark for experimental testing for algorithms that solve this problem. We conclude that, given limited time, a Local Search algorithm starting with a solution found by the Argmax-Product algorithm reaches, on average, better results on the tested datasets. |
|---|