Full text
| |
| Author(s): |
Aline Cristina de Souza
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Matemática, Estatística e Computação Científica |
| Defense date: | 2018-10-11 |
| Examining board members: |
Marcos Eduardo Ribeiro do Valle Mesquita;
Fernando Antonio Campos Gomide;
Peter Sussner;
Marley Maria Bernardes Rebuzzi Vellasco;
Sandra Aparecida Sandri
|
| Advisor: | Marcos Eduardo Ribeiro do Valle Mesquita |
| Abstract | |
Associative memories are mathematical models whose main objective is to store and recall information by association. Such models are designed for the storage a finite set of pairs, called fundamental memory set, and they must present certain noise tolerance, that is, they should be able to retrieve a stored information even from an incomplete or corrupted version of a memorized item. The recurrent correlation associative memories (RCAMs), introduced by Chiueh and Goodman, present large storage capacity and excellent noise tolerance. However, RCAMs are designed to store and retrieve bipolar patterns. The generalized recurrent exponential fuzzy associative memories (GRE-FAMs) can be seen as a generalized version of RCAMs capable of storing and retrieving fuzzy sets. In this thesis, we introduce the generalized exponential bidirectional fuzzy associative memories (GEB-FAMs), an extension of GRE-FAMs to the heteroassociative case. Since GEB-FAMs are based on a similarity measure, we conducted a study of several measures from the literature, including the cardinality based similarity measure and the structural similarity index (SSIM). Furthermore, we show that GEB-FAMs exhibit optimal storage capacity and we present a characterization of the output of a single-step GEB-FAM when one of its parameters tends to infinity. However, in computational experiments, good results were obtained by a single-step GEB-FAM with parameter values in the interval [1,10]. As the dynamics of the GEB-FAMs is still not fully understood, this fact led to a more detailed study of the single-step GEB-FAMs, refered to as fuzzy kernel associative memories (fuzzy-KAMs). We interpret this model by using a fuzzy kernel and we propose to adjust its parameter by using the concept of entropy. Also, we present two approaches to pattern classification using the fuzzy-KAMs. Finally, we describe computational experiments used to evaluate the performance of such approaches in classification and face recognition problems. In most of the experiments performed, in both types of problems, the classifiers defined based on the proposed approaches obtained satisfactory and competitive performance with those obtained by other models from the literature, which shows the versatility of such approaches (AU) | |
| FAPESP's process: | 15/00745-1 - A Study on Recurrent Exponential Fuzzy Associative Memories, Their Generalizations, and Applications |
| Grantee: | Aline Cristina de Souza |
| Support Opportunities: | Scholarships in Brazil - Doctorate |