CONSTRUCTIVE NETWORKS Constructive networks have increasingly become a popular alternative to the more typical multilayer perceptron network (MLP) trained with the error back-propagation algorithm. The most important characteristic of such networks is the fact that processing units can be gradually included in the network structure during the learning process, avoiding the problem of the prior definition of the number of neurons in the hidden layers necessary to solve each given task. This work investigates some aspects concerning the performance of constructive networks as compared to other architectures such as the MLP, as well as its application to environments which require a continuous modification of the knowledge stored in the network.
Other People Involved:
Juliana Neiva de Gouvea Ribeiro - MSc Student
SPURIOUS PATTERN DETECTION
One very important aspect to take into account when developing neural network models for problems of pattern verification, process control and pattern classification is the ability of the model employed to deal with the ocurrence of patterns that do not belong to any of the classes of patterns used to train the network. This refers to the problem of the detection of spurious or novel patterns (nca.ps.gz). This work aims at the investigation of the fundamental characteristics associated with the problem, the development of neural networks more robust in this sense and the definition of auxiliary mechanisms that make the system at hand more reliable as a whole.
CATASTROPHIC FORGETTING IN NEURAL NETWORKS
Catastophic forgetting refers to the inability of some neural network models to maintain old pattern representations stored in the network in the presence of new input-output associations. Previous pattern mappings are lost when new associations are stored in the network unless the old mappings are continuosly re-learned. The objective of this project is to investigate solutions to the problem through the distribution control of the pattern representations in the network (semi-distributed representations, for example) or through the development of neural structures inherently more efficient in this respect.
NEURAL NETWORKS FOR FINANCE AND BUSINESS 
Neural networks have been increasingly used to solving problems in finance and business. A recent article in The Economist points out that applications and sales involving the use of neural networks in this area will reach 1 Billion Dollars in 1997. UK High Street Banks, Visa, Mastercard, American Express, USA Bank, Britvic Soft Drinks are examples of instituitions that are investing in neural networks as an efficient solution of information technology for their problems. Credit analysis, fraud detection, currency exchange rate prediction, stock selection, prediction of sales and consumer behaviour are only a small number of applications where neural networks have already been successfully employed. This project explores the potential use of neural networks in a variety of applications in finance and business.
Other People Involved:
Paulo J. L. Adeodato - Visiting Professor
Domingos Sávio M.P. Monteiro - MSc Student
SISTEMA NEURAL PARA ANÁLISE DE CRÉDITO
Prototipação de módulo de scoring de clientes para um sistema de Análise de
Crédito, utilizando a tecnologia de Redes Neurais, junto ao
Centro de Estudos e
Sistemas Avançados do Recife. Este projeto está sendo desenvolvido para o Grupo Bompreço S.A.
e encontra-se atualmente em andamento, já tendo sido finalizado um protótipo do
sistema.
Other People Involved:
Paulo J. L. Adeodato - Visiting Professor
Domingos Sávio M.P. Monteiro - MSc Student
SAPRI - AUTOMATIC ACQUISITION,
PRE-PROCESSING AND RECOGNITION OF RADAR IMAGES
The SAPRI project is a joint work supported by ProTeM/CNpQ - Phase 3 (Multi-institutional Programme in Computer Science) involving the Department of Computer Science - UFPE, the Brazilian Navy Research Institute (IPqM) and four other academic institutions (USP, UNICAMP, UFRGS and UFG) whose objective is the development of a system based on advanced methods of computational intelligence such as neural networks, pattern recognition and image processing techniques for the automatic manipulation and interpretation of radar images. This system, composed of five fundamental modules -- aquisition, processing, recognition, tactical analysis, and a knowledge base of nautical documents -- will be employed by the Brazilian Navy in the tasks of aero-naval traffic control (e.g. automatic target detection) and continuous monitoring and verification of navigation rules.
Other People Involved:
Edson C. B. Carvalho Filho - Associate Professor
Paulo J. L. Adeodato - Visiting Professor
Alejandro Frery Orgambide - Associate Professor
Teresa B. Ludermir - Associate Professor
Eduardo de Aguiar Sodré - Research Assistant
Fausto de Oliveira Queiroz - Undergraduate Student
SISTEMAS DE APOIO À DECISÃO, PREVISÃO
e CONTROLE
Em muitas aplicações do mundo real, ambientes precisam ser continuamente monitorados
a fim de diferenciar entre condições normais e anormais de operação de um
sistema. Este tipo de situação é particularmente importante em aplicações
onde a questão de segurança é um fator crítico como em verificação
da autenticidade de padrões, detecção e prevenção automática de falhas,
controle de processos industriais e problemas de previsão em geral.
Dois trabalhos desenvolvidos
no momento correspondem a uma dissertação de mestrado
na previsão de cargas em sistemas elétricos e outra de aplicação de redes neurais
em recursos
hídricos (previsão de vazões médias nos reservatórios) de alunos funcionários
da Companhia Hidro-Elétrica do São Francisco
.
Other People Involved:
Arlindo Lins de Araújo Junior - MSc Student
SISTEMA PARA MONITORAMENTO E
ANALISE DE SINAIS ELETROCARDIOGRAFICOS EM ANIMAIS DE PEQUENO PORTE
A ser descrito.
Other People Involved:
Junior Eduardo C. Valoes - MSc Student
REDES NEURAIS PARA CLASSIFICAÇÃO e
VERIFICAÇÃO DE PADRÕES
Este projeto investiga o emprego de redes neurais contíuas (multilayer perceptron, radial basis function, self-organizing maps) em aplicações práticas como reconhecimento de caracteres manuscritos, verificação de assinaturas e diagnose médica. Há também o interesse pelo desenvolvimento de modelos no problema de classificação de impressões digitais. Redes neurais se apresenta como uma abordagem promissora em aplicações desta natureza e o seu emprego em problemas de automação bancária, em sistemas para a detecção de fraudes em cartão de crédito e em sistemas de segurança aparece como uma realidade hoje em dia, já sendo encontrados no mercado exemplos de produtos a nível comercial.
Other People Involved:
Juliana Neiva Gouvea Ribeiro - MSc Student
THE SURFACE ANIMATOR PROGRAM (NeuroSAN) This work consists of the development of a software tool in C for X Windows to be employed as an auxiliary element in the analysis and monitoring of neural network behaviour in pattern recognition and classification tasks. This represents a graphical interface for the visualization in the 3-D space of the decision regions generated by the network during the process of learning. This software tool, the Surface Animator Program, can be used as a powerfull aid in the evaluation of network characteristics such as the evolution time of the learning phase or the observation of the class separation surfaces gradually defined in the pattern space.
Other People Involved:
Flávia Cristina T. do Amaral - Undergraduate Student
Patricia Gouveia Ramos - Undergraduate Student
