Título: IGMN: A New Neural Network Model Based on Gaussian Mixture Models for Incremental Function Approximation and Clustering
Linha de Pesquisa: Inteligência Artificial

Data: 08/10/2010
Horário: 14h
Local: Auditório José M. V. Castilho, prédio 43424 (72)

Banca Examinadora:
Prof. Dr. Gerson Zaverucha (UFRJ)
Prof. Dr. Mauro Roisenberg (UFSC)
Prof. Dr. Luís da Cunha Lamb (UFRGS)

Presidente da Banca: Prof. Dr. Paulo Martins Engel

Resumo: This work proposes a new artificial neural network (ANN) model, called IGMN (for Incremental Gaussian Mixture Network), specially designed for incremental function approximation and clustering tasks. This new neural network model, which is based on multivariate Gaussian Mixture Models, has the following advantages over other neural network models: (i) it can learn instantaneously from data flows considering each training pattern just once (each training point can be immediately used and discarded); (ii) IGMN can learn continuously and perpetually, i.e., there are no separate phases for training and recalling; (iii) the neural network topology is defined automatically and incrementally (new units are added whenever is necessary); (iv) the stability-plasticity dilemma is tackled by a minimum likelihood criterion which decides if a new processing unit must be created; and (v) the learning algorithm is computationally efficient, which allows its use in real time applications. All these advantages make possible to use IGMN in several state-of-art tasks like incremental concept formation, reinforcement learning and robotic mapping. This work also presents several experiments which show that IGMN is an useful tool for incremental function approximation and clustering. Moreover, the IGMN incremental learning algorithm opens new directions where artificial neural networks can be used successfully and efficiently.

Palavras chave: Machine Learning, Artificial Neural Networks, Incremental Learning, Bayesian Methods, Gaussian Mixture Models, Function Approximation, Clustering, Concept Formation, Reinforcement Learning, Autonomous Mobile Robots.