Catálogo de publicaciones - libros

In the post-genomic era, bioinformatics methods play a central role in understanding vast amounts of biological data. Due to their ability to find arbitrarily complex patterns within these data, neural networks play a unique, exciting and pivotal role in areas as diverse as protein structure and function prediction. This paper presents a critical overview of recent advances in bioinformatics which have utilised neural network methods.

According to the Support Vector Machine algorithm, the task of classification depends on a subset of the original data-set, which is the set of Support Vectors (SVs). They are the only information needed to compute the discriminating function between the classes and, therefore, to classify new data. Since both the computational complexity and the memory requirements of the algorithm depend on the number of SVs, this property is very appealing from the point of view of hardware implementations. For this reason, many researchers have proposed new methods to reduce the number of SVs, even at the expenses of a larger error rate. We propose in this work a method which aims at finding a single point per each class, called archetype, which allows to reconstruct the classifier found by the SVM algorithm, without suffering any classification rate loss. The method is also extended to the case of non-linear classification by finding an approximation of the archetypes in the input space, which maintain the ability to classify the data with a moderate increase of the error rate.

In this paper we describe a new method, based on a genetic algorithm, for generating good (in terms of minimum distance) linear block error-correcting codes. We offer a detailed description of the algorithm, with particular regard to the genetic operators (selection, mutation and crossover) which have been specifically adapted to the problem. Preliminary experimental results indicate that the method can be very effective, especially in terms of fast production of good sub-optimal codes.

This paper describes the design of a digital architecture suitable for the classification of large quantities of measurement data by means of a method based on the Support Vector Machines (SVMs). The proposed approach can be applied for solving general inverse modeling problems and for processing complex measurement data requiring real-time processing, possibly in a distributed mode over a number of physically small and geographically separated ‘computational nodes’. A problem of nonlinear channel equalization and a classification task from high energy physics are presented as discussed as two case studies for which the ability of achieving real-time processing is of paramount importance. The performance of such architectures is then analyzed in terms of its speed of execution, occupancy of the hardware modules available in a Virtex II FPGA chip, and classification error.

The application of methods of statistical learning to the identification of the partons from which hadronic jets originate is investigated using simulated jets in the CDF detector with the ultimate objective of applying them at the trigger level. Using only jet-related properties, it appears to be raltively easy to distinguish between jets originating from gluons and those originating from quarks in an energy-independent manner. Distinguishing between quark flavours is more difficult and will require inclusion of other variables.

The liquid state machine is a novel computation paradigm based on the transient dynamics of recurrent neural circuitry. In this paper it is shown that this systems can be used to recognize complex stimuli composed by non-periodic signals and to classify them in a very short time. Even if the network is trained over a segment of the signal the classification task is completed in a time interval significantly shorter than the time-window used for the training. Stimuli composed by many complex signals are recognized and classified even if some signals are absent.

The goal of this project is to realize an enhanced data mining system which performs intelligent processing on data received from sensorial agents in a very flexible manner with reusability prospective. The project is implemented through a “digital core” constituted of a FPGA, a microcontroller and several memory blocks which co-operate to the computation. The FPGA is programmed in VHDL to implement the data mining process. The data mining system is composed of a sophisticated statistical non parametric part and a recurrent artificial neural network. The core was written in a recursive manner to permit the reconfigurability of the network and its reusability to all the systems which can be modeled through a similar system.

High Energy Physics (HEP) experiments require discrimination of a few interesting events among a huge number of background events generated during an experiment. Hierarchical triggering hardware architectures are needed to perform this tasks in real-time. In this paper three neural network models are studied as possible candidate for such systems. A modified Multi-Layer Perceptron (MLP) architecture and a EαNet architecture are compared against a traditional MLP. Test error below 25% is archived by all architectures in two different simulation strategies. EαNet performance are 1 to 2%better on test error with respect to the other two architectures using the smaller network topology. The design of a digital implementation of the proposed neural network is also outlined.

In this paper we propose the adoption of the Random Neural Network Model for the solution of the dynamic version of the Steiner Tree Problem in Networks (SPN). The Random Neural Network (RNN) is adopted as a heuristic capable of improving solutions achieved by previously proposed dynamic algorithms. We adapt the RNN model in order to map the network characteristics during a multicast transmission. The proposed methodology is validated by means of extensive experiments.

We present a package for R language containing a set of tools for regression using ensembles of learning machines and for time series forecasting. The package contains implementations of Bagging and Adaboost for regression, and algorithms for computing mutual information, autocorrelation and false nearest neighbors.