Catálogo de publicaciones - libros

This paper describes a new platform for FPGA implementation of the multilayer perceptron (MLP) back propagation algorithm (BP). The three implementation figures of the algorithm are considered. These are the off type implementation, the on chip global implementation and the dynamic reconfiguration of the ANN. To achieve our goal, a design for reuse strategy has been applied. To validate our approach, three case studies are considered using the Virtex-II and Virtex-4 FPGA devices. A comparative study is done and new conclusions are given.

We present a system that integrates a retinomorphic chip into a platform that includes a board with reconfigurable hardware (FPGA device) and a conventional computer in order to evaluate image processing schemes, such as motion detection based on this front-end. We have used an artificial retina that transforms light intensity into spikes and sends them using an event-driven protocol. To set up this development platform, we have built a driver for a board with a FPGA that acts as an interface between the retina and a personal computer in which we store the grabbed spikes. Also, we have developed software modules to test the vision algorithms before programming them in Hardware Description Languages.

A wavelet-based procedure for clustering signals is proposed. It combines an individual signal preprocessing by wavelet denoising, a dimensionality reduction step by wavelet compression and a classical clustering strategy applied to a suitably chosen set of wavelet coefficients. The ability of wavelets to cope with signals of arbitrary or time-dependent regularity as well as to concentrate signal energy in few large coefficients, offers a useful tool to carry out both significant noise reduction and efficient compression. A simulated example and an electrical dataset are considered to illustrate the value of introducing wavelets for clustering such complex data.

This paper presents a methodology for functional data analysis. It consists in extracting a large number of features with maximal content of information and then selecting the appropriate ones through a Mutual Information criterion; next, this reduced set of features is used to build a classifier. The methodology is applied to an industrial problem: the classification of the dynamic properties of elastomeric material characterized by rigidity and hysteresis curves.

We propose an easy-to-use methodology to allocate one of the groups which have been previously built from a complete learning data base, to new individuals. The learning data base contains continuous and categorical variables for each individual. The groups (clusters) are built by using only the continuous variables and described with the help of the categorical ones. For the new individuals, only the categorical variables are available, and it is necessary to define a model which computes the probabilities to belong to each of the clusters, by using only the categorical variables. Then this model provides a decision rule to assign the new individuals and gives an efficient tool to decision-makers.

This tool is shown to be very efficient for customers allocation in consumer clusters for marketing purposes, for example.

Prototype based neural clustering or data mining methods such as the self-organizing map or neural gas constitute intuitive and powerful machine learning tools for a variety of application areas. However, the classical methods are restricted to data embedded in a real vector space and have only limited applicability to noneuclidean data as occurs in, for example, biomedical or symbolic fields. Recently, extensions of unsupervised neural prototype based clustering to dissimilarity data, i.e. data characterized in terms of a dissimilarity matrix only, have been proposed substituting the mean by the so-called generalized median. Thereby, the location of prototypes is chosen within the discrete input space which constitutes a severe limitation in particular for sparse data sets since the prototype flexibility is restricted. Here we present a generalization of median neural gas such that prototypes can be interpreted as mixtures of discrete input locations. We derive a batch optimization scheme based on a corresponding cost function.

The present paper aims at locating the breakings of the integration process of an international system observed during about 50 years in the 19th century. A historical study could link them to special events, which operated as exogenous shocks on this process. The indicator of integration used is the spread between the highest and the lowest among the London, Hamburg and Paris gold-silver prices. Three algorithms are combined to study this integration: a periodization obtained with the SOM algorithm is confronted to the estimation of a two-regime Markov switching model, in order to give an interpretation of the changes of regime; in the same time change-points are identified over the whole period providing a more precise interpretation of the various types of regulation.

We extend the self-organizing map to a supervised fuzzy classification method. On the one hand, this leads to a robust classifier where efficient learning of fuzzy labeled or partially contradictory data is possible. On the other hand, class similarities may be detected. Further, we integrate the parametrized functional metric as data similarity measure into the approach. Parametrization of this functional metric allows relevance learning for efficient classification and feature selection.

This paper contribution is about numerical methods based on interval analysis for approximating sets, and about the application of these methods to vast classes of statistical problems. ’Guaranteed’ means here the inner and outer approximations of the sets of interest are obtained, which can be made as precise as desired, at the cost of increasing the computational effort. It thus becomes possible to archieve tasks still thought by many to be out of the reach of numerical methods, such as finding all solutions of sets of non-linear equations and inequalities or a global optimizer of possible multi-modal criteria.

Currently, there exist many research areas that produce large multivariable datasets that are difficult to visualize in order to extract useful information. Kohonen self-organizing maps have been used successfully in the visualization and analysis of multidimensional data. In this work, a projection technique that compresses multidimensional datasets into two dimensional space using growing self-organizing maps is described. With this embedding scheme, traditional Kohonen visualization methods have been implemented using growing cell structures networks. New graphical map displays have been compared with Kohonen graphs using two groups of simulated data and one group of real multidimensional data selected from a satellite scene.