It was more than 50 years ago when John Von Neumann introduced his groundbreaking work on
self-organization in cellular automata and on the intimate relationship between computers and the
brain [24]. In the past decades, in particular since the rebirth and explosive development
of the field in the mid 80s, neural network approaches have demonstrated their excellent
potential to support fundamental theoretical and practical research towards new generations
of artificial intelligence and intelligent computing. Practical models uphold a variety of
applications in engineering and computer science, and also in many other fields such as finance,
bioinformatics and medicine. In most of these areas, some examples are given for the methodologies
(and applications) the author (in collaboration with other researchers) has worked with.
Multi-Layer Perceptron MLP
Intelligent signal processing is a major field where neural networks have clearly demonstrated their
advantages over traditional statistical techniques. MLP is capable of adaptively approximating the
function of any linear or nonlinear filter, and thus is very competitive when applied to data
representation such as the extraction of efficient discriminative features, filtering, and enhancement of
signals. Its capability of function approximation motivated our initial interest to use neural
networks for problem solving in engineering. In [14] a theorem based on the seminal work of
Kolmogorov for function approximation is proven. In addition, discussion of its impact
to neural networks in the context of existence (and constructiveness) is provided. MLPs
have been used in varieties of intelligent signal processing tasks, such as internal model
control of a lime kiln [34, 33], fault detection and diagnosis in rotary lime kiln [35, 36],
or in parts plastic industry [26, 22], For example, in data presentation, MLP has been
successfully used in plastics image segmentation, feature location and parts defect detection in
an injection molding machine [21]. An approach for finding the architecture of an MLP
looking at the geometry of a given classification task was proposed in [36]. Pursuing a more
algorithmic approach, a new architecture of an MLP-based neural network incorporating both
hybrid learning and multi-topology has been developed in [23]. A new class designated
Multiple Feed-Forward (MFF) networks, and the new gradient-based learning algorithm,
Multiple Back-Propagation (MBP), are both proposed and analyzed. This new architecture
(and algorithm) has a number of advantages over classical ones in that it can be trained to
be noise-robust, and thus is more suitable for real-world environments in manufacturing
industry [22]. Moreover, the methodology is very valuable in the pre-processing and feature
extraction of engineering (and biomedical) data, and has been used in solving difficult
signal processing and pattern recognition problems. Across the very successful applications
we include information management [32], credit risk prediction [40] and medicine [37].
Modular Neural Networks
The feasibility of applying neural networks, specifically modular constructive neural networks, to
mobile robot navigation was addressed in [48]. Due to their adaptive and generalization capabilities
they are a reasonable alternative to more traditional approaches. The problem, in its basic form,
consists of defining and executing a trajectory to a pre-defined goal while avoiding all obstacles, in an
unknown environment. Some crucial issues arise when trying to solve this problem, such
as an overflow of sensorial information and conflicting objectives. From this analysis, the
necessity of introducing the concept of modularity has arisen, as a way to circumvent the
existing conflicts. An original modular architecture instead of a monolithic approach is
presented in [45] that effectively combines all the information available and navigates the
robot through an unknown environment towards a goal position. In addition to modularity
another concept is used - constructiveness - to find the best possible architecture [46].
The proposed architecture is successfully tested with the NOMAD 200TM mobile robot.
Associative and Hopfield Neural Networks
Hopfield neural networks are proposed in [6] as an alternative to Dijkstra’s shortest path algorithm for routing in Quality of Service (QoS) data networks. The proposed solution extends the traditional single-layer recurrent Hopfield architecture introducing a two-layer architecture that automatically guarantees an entire set of constraints held by any valid solution to the shortest path problem. This new method addresses some of the limitations of previous solutions, in particular the lack of reliability in what concerns successful and valid convergence. Experimental results show that an improvement in successful convergence can be achieved in certain classes of graphs. Additionally, computation performance is also improved at the expense of slightly worse results.