Ebook: Neural Nets WIRN11

This book is a collection of selected papers from the 21st edition of the WIRN workshop.

Along with the regular sessions, the annual meeting of the Italian Neural Networks Society (SIREN) customarily includes special sessions where conference participants discuss and analyze the latest challenges in the field of neural networks and related topics. Similarly, in the tradition of the workshop, the open-minded and friendly style of the discussions attract the contribution of researchers from many different countries, as well as a solid community of historical members who have grown the society through its 21 years.

All the authors have presented original results, and peer reviewers worked generously to enhance the texts by suggesting improvements and remedying possible lapses.

As for the content; the regular session papers are divided as usual into three groups: Applications, Models and specific Signal Processing implementations. Then come the contributions to the following special sessions: Autonomous Machine Learning, Neuromorphic Engineering and Models of Behavior for Human-Machine Interaction. The first special session was organized in cooperation with INNS SIG AML (International Neural Networks Society – Special Interest Groups on Autonomous Machine Learning); the last constitutes a special activity of the European COST (European Cooperation in Science and Technology) Action 2012 on Cross-Modal Analysis of Verbal and Nonverbal Communication.

The Editors would like to take this opportunity to thank the invited speakers and contributors as well as the referees for their efforts in the realization of this volume, and look forward to meeting them all again next year.

Proteomics has gained a wide interest in the last decade since it involves the comparative study of protein expressions to identify bio-markers for early diagnosis of unpredictable, and serious, pathologies. The most powerful techniques for protein investigation compare 2D gel electrophoresys images that represent the protein composition of healthy and diseased tissues. Nevertheless, this analysis is problematic since gel images are affected by high noise levels and they are distorted, so that the same protein spot has different locations on different gels. Furthermore, the acquisition of a statistically significant sample of gels from a unique laboratory is problematic due to ethical problems, to the rarity of certain diseases, and to the fact that the process of gel electrophoresys is time consuming and costly. However, a great deal of information is present in the scientific literature in the form of images reporting 2D gels acquired for different experiments, we have developed a framework to compare annotated 2D gel images extracted from state of the art, and publicly available papers. The system has been assessed by performing the comparative analysis of the Haptoglobin; although the analyzed images are much more noisy and distorted than their sources the system achieves promising results.

In this paper we show how combining fuzzy sets and reinforcement learning a winning agent can be created for the popular Pac-man game. Key elements are the classification of the state into a few fuzzy classes that makes the problem manageable. Pac-man policy is defined in terms of fuzzy actions that are defuzzified to produce the actual Pac-man move. A few heuristics allow making the Pac-man strategy very similar to the Human one. Ghosts agents, on their side, are endowed also with fuzzy behavior inspired by original design strategy. Performance of this Pac-man is shown to be superior to those of other AI-based Pac-man described in the literature.

This work presents a methodology to improve soft sensors performances in spatial forecast of environmental parameters. To this aim, we substitute a single soft sensor based on a single neural network with a more complex connectionist system that we call the HyperSensor. HyperSensor is built by a set of soft sensors; each one based on a specific neural model and a gating neural network, which plays the role of a stochastic selector. HyperSensor wraps the best characteristics of different neural network models through the gating network, which selects the best performing soft sensor according to the current input. In other words, HyperSensor is able to independently choose the best instrument of measure to get the best performance.

Graph Echo State Network (GraphESN) is an efficient neural network model that extends the applicability of Reservoir Computing to the processing of graphs. The untrained reservoir encodes input graphs into isomorphic structured states. In this paper we propose a novel and supervised approach for adaptively weight the relevance of the states of the vertices in the input graphs for the output computation in classification tasks. To this aim, local average computations on partitions of the state space, obtained using the Neural Gas algorithm, are combined according to the target information. The effectiveness of the proposed approach is shown on real-world tasks from Cheminformatics.

The measurement of the quality of research has reached nowadays an increasing interest not only for scientific reasons but also for the critical problem of researchers' ranking, due to the lack of grant assignments. The most commonly used approach is based on the so-called h-index, even if the current literature debated a lot about its pros and cons. This paper, after a brief review of the h-index and of alternative models, focuses on the characterization and the implementation of a modified scoring rule approach by means of a fuzzy inference system a là Sugeno.

This work shows empiric evidence about the use of probabilistic sets to rank a set of t-commerce applications, keeping in mind the perceived usefulness to determine the preferences of users. To measure these preferences has been developed a simulated market test with twelve t-commerce designs. Each design represents a combination of rational and emotional variables. The methodology allows rank the preferences of users, measured with an ordinal variable, keeping all information supplied by the sample. A coherent result has been obtained in spite of important uncertain degree in original data.

The last effusive eruption on February 27, 2007 at Stromboli volcano was characterized by the occurrence of a particular typology of seismic events named “hybrids”. During March about 4000 of these signals were recorded, and three main swarms happened: the first one on days 6-8, with more than 1200 events; the second one on day 20, with more than 400 events; and the third one on day 22, with about 600 events. The study of these events and specifically their location is the main purpose of this work because it not only characterizes a particular aspect of the 2007 effusive eruption but at the same time can improve the understanding of the eruptive processes of the volcano. Thus, in order to locate them it was first necessary to group the signals according to their waveform similarity and then apply relative location techniques on individual families. To perform the clustering an unsupervised SOM neural network was used. This technique is capable of working without any “a-priori” information about data distribution and structure. Its results have revealed differences in the families of events recorded during and between the swarms, underlying from a volcanological point different locations or source mechanisms of the involved structures. Moreover, they have shown to be consistent compared to those obtained by applying the Hierarchical Clustering technique. However, in contrast to the latter, the SOM clustering does not critically depend on its parameters and allows for an easier result visualization and interpretation.

In recent years biological processes modeling and simulation have become two key issues in analyzing complex cellular systems. Information about metabolic networks is often incomplete, since a large portion of available data is ignored by its probabilistic nature. The main objective of this work is to investigate metabolic networks behavior in terms of their fault tolerance capabilities as random node removal and high-connectivity-degree node removal aimed at affecting network activities. The paper proposes a software framework, namely CEllDataLaB, based on three techniques to perform the structural and functional analysis of a metabolic network: topological analysis, flux balance analysis and extreme pathways analysis. The degradation of proteins into aminoacids metabolic network has been used to validate the implemented investigations strategies. The performed trials have shown that the node connectivity degrees as well as the node functional role in the network are key issues to evaluate the impact of node deletion on network behavior and activities.

Time-Varying Neural Network (TV-NN) is a novel structure applied to nonstationary system identification tasks. Up to date, there are two main categories of approaches to train TV-NN: Gradient-based and ELM-based. Among the latter, the variants EM-ELM-TV and EM-OB have been recently proposed by the authors to determine the number of hidden nodes and the number of output bases functions automatically, which are important parameters to be preset in standard ELM. The aim of this contribution consists in evaluating the performances of aforementioned ELM-based algorithms in training TV-NNs to identify nonstationary Volterra systems, which are used to model a wide category of nonstationary nonlinear systems. Simulation results show that with polynomial activation function, ELM-based algorithms are able to attain good generalization performances in the addressed identification problem.

In the general framework of kernel machines, the adoption of the hinge loss has become more popular than square loss, also because of computational reasons. Since learning reduces to a linear system of equations, in case of very large tasks for which the number of examples is proportional to the input dimension, the solution of square loss regularization is O(ℓ3), where ℓ is the number of examples, and it has been claimed that learning is unaffordable for large scale problems. However, this is only an upper bound, and in-depth experimental analyses indicate that for linear kernels (or in other cases where the kernel matrix will be sparse or decomposed in a way that is known a priori), regularized least square (RLS) is substantially faster than support vector machine (SVM) both at training and test times. In this paper, we give theoretical results to support those experimental findings by proving that there are conditions under which learning of square loss regularization is Θ(ℓ) even for large input dimensions d for which d ≃ ℓ.

We show how to build an associative memory from a finite list of examples. By means of a fully-blown example, we demostrate how a probabilistic Bayesian factor graph can integrate naturally the discrete information contained in the list with smooth inference.

In this paper, the relation between synchronization and control of chaotic nodes connected through a time–varying network is discussed. In particular, the effects of pinning control on a set of moving chaotic agents are investigated showing that the role of system parameters, like agent density, is critical in order to reach the synchronous behavior and also to control the whole network by pinning a reduced set of agents.

Quality assessment in clustering is a long-standing problem. In this contribution we describe some indexes to measure properties of clusterings, taking advantage of the added flexibility provided by fuzzy paradigms. We first present an approach to evaluate some indicators of quality of an individual clustering, by analyzing the co-association matrix. Then we describe a technique to evaluate the similarity of pairs of clusterings by comparing their respective co-association matrices by means of generalizations of well-known indexes of partition comparisons. Finally, we illustrate how some indexes borrowed from spectral graph theory can be used to evaluate clustering stability and diversity in ensembles of several clusterings.

In this work we propose a new approach to the stability analysis of Random Boolean Networks (RBNs). In particular, we focus on two families of RBNs with k=2, in which only two subsets of canalizing Boolean function are allowed, and we show that the usual measure of RBNs stability – sometimes known as the Derrida parameter (DP) – is similar in the two cases, while their dynamics (e.g. number of attractors, length of cycles, number of frozen nodes) are different. For this reason we have introduced a new measure, that we have called attractor sensitivity (AS), computed in a way similar to DP, but perturbing only the attractors of the networks. It is proven that AS turns out to be different in the two cases analyzed. Finally, we investigate Boolean networks with k=3, tailored to solve the Density Classification Problem, and we show that also in this case the AS describes the system dynamical stability.

In this work we propose a novel geometric clustering algorithm based on the Tensor Voting Framework (TVF). More precisely, we propose the construction of a weighted graph by means of the information diffused by TVF during the vote casting step. This graph, which summarizes informations related to the manifold geometric structure, was used for clustering purposes. To this aim, we applied the well known Dijkstra and Ford Fulkerson algorithms to recursively separate weakly connected graph components.

We performed preliminary tests, comparing our algorithm with that obtained by employing a weighted version of the ε-NN graph. The obtained results on both synthetic and real data show that the proposed technique is promising. To test our algorithm on real datasets, we preprocessed graylevel input images by extracting their edge pixel points.

This paper proposes a novel type of quantum-inspired evolutionary algorithm (QiEA) for numerical optimization inspired by the multiple universes principle of quantum computing, which is based on the concept and principles of quantum computing, such as a quantum bit and superposition of states. Numerical optimization problems are an important field of research with several applications in several areas: industrial plant optimization, data mining and many others, and although being successfully used for solving several optimization problems, evolutionary algorithms still present issues that can reduce their performances when faced with task where the evaluation function is computationally intensive. In order to address those issues the QiEA represent the most recent advance in the field of evolutionary computation. This work present some application about combinatorial and numerical optimization problems.

Feature statistics normalization in the cepstral domain is one of the most performing approaches for robust automatic Speech Recognition (ASR) in noisy acoustic scenarios. According to this approach, feature coefficients are normalized by using suitable linear or nonlinear transformations in order to match the noisy speech statistics to the clean speech one. Histogram Equalization (HEQ) is an effective algorithm belonging to this category. Recently some of the authors have proposed an interesting extension to the HEQ original algorithm, in order to suitably deal with the multichannel audio information coming from multi-microphone sensory activity in far-field acoustic scenarios. In this paper the feature normalization capabilities of the multichannel HEQ technique are further enhanced by introducing the kernel estimation technique and employing the multi-condition training for ASR system parametrization. Computer simulations based on the Aurora 2 database have shown that a significant recognition improvement with respect to the single-channel counterpart and other multi-channel techniques can be achieved confirming the effectiveness of the idea.

Size constrained clustering has been recently proposed to embed “a priori” knowledge in clustering methods. By exploiting the “string property” we propose an exact and efficient algorithm based on dynamic programming techniques to solve size-constrained one-dimensional clustering problems. We show the applicability of the proposed method in a difficult computational biology problem: the prediction of the transcription start sites of genes. The obtained experimental results clearly show the potential of the proposed approach when compared with previously published methods.

Electroencephalographic (EEG) recordings are employed in order to investigate the brain activity in neuropathological subjects, but unfortunately EEG are often contaminated by artifacts, signals that have no-cerebral origin and therefore distort the EEG analysis. We know that entropy measures reflect the degree of order/disorder of the EEG signal, so that is represents a good instrument for artifacts detection. In this paper we propose a multiresolution analysis, based on EEG wavelet processing, to extract cerebral EEG rhythms. The novelty of this paper is to apply the Wavelet Entropy method not only to Shannon Entropy formulation, but also to Rényi Entropy and Tsallis Entropy formulation in order to characterize the functional dynamics of EEG signal.

In this paper a collaborative filter combination for time-series prediction is considered. The basic idea is based on a convex combination of two kernel adaptive filters with different parameters. While the convergence of one filter is fast but not accurate, the convergence of the second one is much more accurate, even if slower. The convex combination of both filters allows to reach good performances in terms of convergence and speed. Some experimental results on the prediction of the Mackey-Glass time-series demonstrate the effectiveness of the proposed approach.

In this paper novel theoretical medical decision support system based on a personalized modeling gene selection method is presented. Identifying a compact set of genes from gene expression data is a critical step in bioinformatics research. Personalized modeling is a recently introduced technique for constructing clinical decision support systems. In this work we have provided a comparative study using the proposed Personalized Modeling based Gene Selection method on two benchmark microarray datasets (Colon cancer and Central Nervous System cancer data) and on a macroarray dataset collected during a GRANT clinical project. The experimental results show that our method is able to identify a small number of informative genes which can lead to reproducible and acceptable predictive performance without expensive computational cost. These genes are of importance for specific groups of people for cancer and other diseases diagnosis and prognosis.