Search

Purpose We present image classifiers based on Dense Convolutional Networks and transfer learning to classify chest X-ray images according to three labels: COVID-19, pneumonia, and normal. Methods We fine-tuned neural networks pretrained on ImageNet and applied a twice transfer learning approach, using NIH ChestX-ray14 dataset as an intermediate step. We also suggested a novelty called output neuron keeping, which changes the twice transfer learning technique. In order to clarify the modus operandi of the models, we used Layer-wise Relevance Propagation (LRP) to generate heatmaps. Results We were able to reach test accuracy of 100% on our test dataset. Twice transfer learning and output neuron keeping showed promising results improving performances, mainly in the beginning of the training process. Although LRP revealed that words on the X-rays can influence the networks’ predictions, we discovered this had only a very small effect on accuracy. Conclusion Although clinical studies and larger datasets are still needed to further ensure good generalization, the state-of-the-art performances we achieved show that, with the help of artificial intelligence, chest X-rays can become a cheap and accurate auxiliary method for COVID-19 diagnosis. Heatmaps generated by LRP improve the interpretability of the deep neural networks and indicate an analytical path for future research on diagnosis. Twice transfer learning with output neuron keeping improved DNN performance.

Motor imagery (MI) has been increasingly studied for neurorehabilitation purposes. However, issues such as large intra- and inter-subject variability still limit its practical, and more clinical, applications. Seeking appropriate features for MI-neuromodulation training is thus a crucial step. This work presents a protocol that selects features related to the MI mental patterns maximizing consistency (i.e., minimizing variability) across two recordings in different days. We apply our methodology to 3 healthy adults and 3 children with cerebral palsy, illustrating its feasibility for MI training protocols.

Objective We used deep convolutional neural networks (DCNNs) to classify electroencephalography (EEG) signals in a steady-state visually evoked potentials (SSVEP) based single-channel brain–computer interface (BCI), which does not require calibration on the user. Methods EEG signals were converted to spectrograms and served as input to train DCNNs using the transfer learning technique. We also modified and applied a data augmentation method, SpecAugment, generally employed for speech recognition. Furthermore, for comparison purposes, we classified the SSVEP dataset using support-vector machines (SVMs) and filter bank canonical correlation analysis (FBCCA). Results Excluding the evaluated user's data from the fine-tuning process, we reached 82.2% mean test accuracy and 0.825 mean F1-Score on 35 subjects from an open dataset, using a small data length (0.5 s), only one electrode (Oz) and the DCNN with transfer learning, window slicing (WS) and SpecAugment's time masks. Conclusion The DCNN results surpassed SVM and FBCCA performances, using a single electrode and a small data length. Transfer learning provided minimal accuracy change, but made training faster. SpecAugment created a small performance improvement and was successfully combined with WS, yielding higher accuracies. Significance We present a new methodology to solve the problem of SSVEP classification using DCNNs. We also modified a speech recognition data augmentation technique and applied it to the context of BCIs. The presented methodology surpassed performances obtained with FBCCA and SVMs (more traditional SSVEP classification methods) in BCIs with small data lengths and one electrode. This type of BCI can be used to develop small and fast systems.

Self-modulation of brain signals related to motor imagery (MI) activity has been increasingly indicated to be beneficial for motor rehabilitation protocols. Appropriately incorporating the MI strategy requires ensuring the active involvement of the central nervous system. Under this context, seeking MIrelated features to be easily modulated by the user and that are recurrent between distinct MI sessions is a necessary and challenging task. This paper proposes a criterion to maximize the consistency of features from the electroencephalography signal between recording sessions performed at different days. To achieve this, we explore three distinct approaches to rank features related to MI’s underlying mental patterns - the event-related desynchronization (ERD) score, the ERD occurrence (EO) and the Fisher ratio- applied to a database of 3 healthy adults and 4 children with cerebral palsy. Overall, the healthy adults presented more consistent features for F-, FC- and C-labeled electrodes and the lower frequency bands for the EO and ERD score, and higher ranges for the Fisher ratio. For the children, results showed more variability but emphasis on specific spectral and spatial locations could also be observed individually. Hence, we claim that distinct ranking metrics may reflect different aspects of MI, but each one may be best for specific types of applications and subjects.

Brain-computer interfaces (BCI) based on steady-state visually evoked potentials (SSVEP) have been increasingly used in different applications, ranging from entertainment to rehabilitation. Filtering techniques are crucial to detect the SSVEP response since they can increase the accuracy of the system. Here, we present an analysis of a space-time filter based on the Minimum Variance Distortionless Response (MVDR). We have compared the performance of a BCI-SSVEP using the MVDR filter to other classical approaches: Common Average Reference (CAR) and Canonical Correlation Analysis (CCA). Moreover, we combined the CAR and MVDR techniques, totalling four filtering scenarios. Feature extraction was performed using Welch periodogram, Fast Fourier transform, and CCA (as extractor) with one and two harmonics. Feature selection was performed by forward wrappers, and a linear classifier was employed for discrimination. The main analyses were carried out over a database of ten volunteers, considering two cases: four and six visual stimuli. The results show that the BCI-SSVEP using the MVDR filter achieves the best performance among the analysed scenarios. Interestingly, the system’s accuracy using the MVDR filter is practically constant even when the number of visual stimuli was increased, whereas degradation was observed for the other techniques.

Agradecimentos: This work was supported by the São Paulo Research Foundation (FAPESP, Brazil – Grant 2013/07559-3), the Studies and Projects Funding Agency(FINEP, Brazil), the National Council of Scientific and Technological Development (CNPq, Brazil), and the Coordination for the Improvement of Higher Education Personnel (CAPES, Brazil) Abstract: The study of motor imagery (MI) has been a subject of great interest within the brain-computer interface (BCI) community. Several approaches have been proposed to solve the problem of classifying cerebral responses due to MI, mostly based on the power spectral density of the mu and beta bands; however, no optimum manner of proceeding through the fundamental steps of a MI-BCI has yet been established. In this work, we explored a relatively novel approach regarding feature generation for a MI-BCI by assuming that functional connectivity patterns of the brain are altered during hand MI. We modelled interactions among EEG electrodes by a graph, extracted metrics from it during left and right hand MI from eight subjects and classified the signals using commonly employed techniques in the BCI community (LDA and SVM). We also compared this approach to the more established method of using the signal power spectral density as the classifier features. With the graph method, we confirmed that only specific electrodes provide relevant information for data classification. A first approach provided maximum average classification rates across all subjects for the graph method of 86% for the mu band and 87% for the beta band. For the PSD method, average rates were of 98% and 99% for the mu and beta bands, respectively. However, a much larger number of features was needed: (130 44) and (273 89) for the mu and beta bands, respectively. Aiming to reproduce these rates using the graph method, pairwise inputs combinations of graph metrics were tested. They proved to be sufficient to obtain essentially the same classification accuracy rates, but with a considerably smaller number of features - about 60 features, for both bands. We thus conclude that the graph method is a feasible option for classification of hand MI signals FUNDAÇÃO DE AMPARO À PESQUISA DO ESTADO DE SÃO PAULO - FAPESP FINANCIADORA DE ESTUDOS E PROJETOS - FINEP CONSELHO NACIONAL DE DESENVOLVIMENTO CIENTÍFICO E TECNOLÓGICO - CNPQ COORDENAÇÃO DE APERFEIÇOAMENTO DE PESSOAL DE NÍVEL SUPERIOR - CAPES Fechado

Agradecimentos: The authors would like to thank FAPESP (Processes n. 2014/27108-9 and 2015/16325-1) and CNPq (Processes n. 311786/2014-6 and 305621/2015-7) for the financial support Abstract: Several problems in signal processing are addressed by expert systems which take into account a set of priors on the sought signals and systems. For instance, blind source separation is often tackled by means of a mono-objective formulation which relies on a separation criterion associated with a given property of the sought signals (sources). However, in many practical situations, there are more than one property to be exploited and, as a consequence, a set of separation criteria may be used to recover the original signals. In this context, this paper addresses the separation problem by means of an approach based on multi-objective optimization. Differently from the existing methods, which provide only one estimate for the original signals, our proposal leads to a set of solutions that can be utilized by the system user to take his/her decision. Results obtained through numerical experiments over a set of biomedical signals highlight the viability of the proposed approach, which provides estimations closer to the mean squared error solutions compared to the ones achieved via a mono-objective formulation. Moreover, since our proposal is quite general, this work also contributes to encourage future researches to develop expert systems that exploit the multi-objective formulation in different source separation problems CONSELHO NACIONAL DE DESENVOLVIMENTO CIENTÍFICO E TECNOLÓGICO - CNPQ FUNDAÇÃO DE AMPARO À PESQUISA DO ESTADO DE SÃO PAULO - FAPESP Fechado

Deep neural networks are becoming the state-of-the-art in the field of machine learning, especially because of their notable performance in a variety of current problems. Among these algorithms we can cite convolutional neural networks as one of the most powerful models for image classification problems. This research aimed to perform a theoretical study on deep neural networks, as well as to implement a convolutional neural network applied to the problem of handwritten digits classification.

In this work, a Brain-Computer Interface based on motor imagery was implemented focusing on the signal processing stage. Methods based on machine learning were used along with feature selection techniques on the treatment and classification of data.

A Brain Computer Interface (BCI) provides a channel of communication and control amid the human brain and a computer or electronic device. In this paper we describe the implementation of all the main steps involved in the operation of a BCI system based on Steady State Visually Evoked Potential (SSVEP) paradigm. Also, we present a comparative study about the performance of our BCI employing two feature extraction techniques (Welch’s Method and Discrete Fourier Transform), and considering the signal analysis window of two sizes (4 and 6 s). The results guided the conception of an online BCI-SSVEP. Our system was successfully controlled by two volunteers, who completed the task of moving a foam block with a robotic manipulator.

Brain connectivity analysis via complex networks has been widely applied to elucidate functional aspects related to brain diseases, such as Alzheimer and Parkinson, and, more recently, to investigations concerning the functional organization of brain regions under motor imagery in brain computer interfaces (BCIs). Therefore, this work seeks to investigate the classification performance of steady-state visually evoked potential (SSVEP) brain-computer interfaces based on functional connectivity. Two different approaches were chosen for extracting functional connectivity and estimating the adjacency matrix from SSVEP-EEG signals: classical Pearson correlation and a new proposal based on Space-Time recurrence counting. These strategies were followed by graph feature evaluation (clustering coefficient, degree, betweenness and eigenvalue centralities), feature selection via Davies-Bouldin index and classification using a least squares classifier for 15 subjects in a 4-command SSVEP-BCI system. For comparison, we also employed a classical spectral feature extraction approach based on the fast Fourier transform (FFT). It was observed that it is possible to separate the classes with a mean accuracy of 0.56 for Pearson and 0.61 for the STR framework, with the clustering coefficient and the eigenvector centrality being the best attributes for these scenarios, respectively. Nonetheless, classical FFT-based feature extraction obtained the best decoding performance.

This work presents a classification performance comparison between different frameworks for functional connectivity evaluation and complex network feature extraction aiming to distinguish motor imagery classes in electroencephalography (EEG)-based brain-computer interfaces (BCIs). The analysis was performed in two online datasets: (1) a classical benchmark-the BCI competition IV dataset 2a-allowing a comparison with a representative set of strategies previously employed in this BCI paradigm and (2) a statistically representative dataset for signal processing technique comparisons over 52 subjects. Besides exploring three classical similarity measures-Pearson correlation, Spearman correlation, and mean phase coherence-this work also proposes a recurrence-based alternative for estimating EEG brain functional connectivity, which takes into account the recurrence density between pairwise electrodes over a time window. These strategies were followed by graph feature evaluation considering clustering coefficient, degree, betweenness centrality, and eigenvector centrality. The features were selected by Fisher's discriminating ratio and classification was performed by a least squares classifier in agreement with classical and online BCI processing strategies. The results revealed that the recurrence-based approach for functional connectivity evaluation was significantly better than the other frameworks, which is probably associated with the use of higher order statistics underlying the electrode joint probability estimation and a higher capability of capturing nonlinear inter-relations. There were no significant differences in performance among the evaluated graph features, but the eigenvector centrality was the best feature regarding processing time. Finally, the best ranked graph-based attributes were found in classical EEG motor cortex positions for the subjects with best performances, relating functional organization and motor activity. Graphical Abstract Evaluating functional connectivity based on Space-Time Recurrence Counting for motor imagery classification in brain-computer interfaces. Recurrences are evaluated between electrodes over a time window, and, after a density threshold, the electrodes adjacency matrix is stablish, leading to a graph. Graph-based topological measures are used for motor imagery classification.

Machine learning has been increasingly used in current days. Great improvements, especially in deep neural networks, helped to boost the achievable performance in computer vision and signal processing applications. Although different techniques were applied for deep architectures, the frequency domain has not been thoroughly explored in this field. In this context, this paper presents a new method for extracting discriminative features according to the Fourier analysis. The proposed frequency extractor layer can be combined with deep architectures in order to improve image classification. Computational experiments were performed on face liveness detection problem, yielding better results than those presented in the literature for the grandtest protocol of Replay-Attack Database. This paper also aims to raise the discussion on how frequency domain layers can be used in deep architectures to further improve the network performance.

Hands motor imagery (MI) has been reported to alter synchronization patterns amongst neurons, yielding variations in the mu and beta bands' power spectral density (PSD) of the electroencephalography (EEG) signal. These alterations have been used in the field of brain-computer interfaces (BCI); in an attempt to assign distinct MI tasks to commands of such a system. Recent studies have highlighted that inforniation may be missing if knowledge about brain functional connectivity is not considered. In this work, we modeled the brain as a graph in which each EEG electrode represents a node. Our goal was to understand if there exists any I near correlation between variations in the synchronization patterns that is, variations in the PSD of mu and beta bands induced by MI and alterations in the corresponding functional networks. Moreover, we (I) explored the feasibility of using functional connectivity parameters as features fora classifier in the context of an MI-BCI; (2) investigated three different types of feature selection (FS) techniques; and (3) compared our approach to a more traditional method using the signal PSD as classifier inputs. Ten healthy subjects participated in this study. We observed significant correlations (p < 0.05) with values ranging from 0.4 to 0.9 between PSD variations and functional network alterations for some electrodes, prominently in the beta band. The PSD method performed better for data classification, with mean accuracies of (90 +/- 8)% and (87 +/- 7)% for the mu and beta band, respectively, versus (83 +/- 8)% and (83 +/- 7)% for the same bands for the graph method. Moreover, the number of features for the graph method was considerably larger. However, results for both methods were relatively close, and even overlapped when the uncertainties of the accuracy rates were considered. Further investigation regarding a careful exploration of other graph metrics may provide better alternatives FUNDAÇÃO DE AMPARO À PESQUISA DO ESTADO DE SÃO PAULO - FAPESP FINANCIADORA DE ESTUDOS E PROJETOS - FINEP CONSELHO NACIONAL DE DESENVOLVIMENTO CIENTÍFICO E TECNOLÓGICO - CNPQ COORDENAÇÃO DE APERFEIÇOAMENTO DE PESSOAL DE NÍVEL SUPERIOR - CAPES Aberto

Echo state networks (ESNs) are recurrent structures that give rise to an interesting trade-off between achievable performance and tractability. This is a consequence of the fact that the key element of these networks – the recurrent intermediate layer known as dynamical reservoir – is not, as a rule, subject to supervised training, which is restricted to the linear output layer, also termed as readout. This trade-off, aside from being of theoretical significance, establishes ESNs as most attractive tools for both online and offline information processing. There are two key aspects to be taken into account in the ESN design: (i) the unsupervised definition of the synaptic weights of the reservoir and (ii) the definition of the structure and of the training strategy associated with the readout. This work is concerned with the first of these aspects: it proposes novel strategies for ESN reservoir design based on the theoretical framework built by Kohonen׳s classical works on self-organization – which includes the notions of short-range positive feedback and lateral inhibition – and also on the related and more recent notion of neural gas. It is shown, with the aid of a representative set of simulation results, that the proposed methodologies are capable of leading to significant performance improvements in the context of relevant information processing tasks – channel equalization and chaotic time series prediction – particularly when the input data suits well a cluster-based profile.

International audience; In the context of Post-Nonlinear (PNL) mixtures, source separation can be performed in a two-stage approach, which encompasses a nonlinear and a linear compensation part. In the former part, however , it is usually required the knowledge of the distribution type for all sources, what may be difficult to attend. In view of this, in this work, we propose a less restrictive approach, in which it is required the knowledge of a single source distribution – here, chosen to be a colored Gaussian. The other sources are only required to present a time structure. The method combines, in a joint-based approach, the use of the second-order statistics (SOS) and the matching of distributions, which shows to be less costly than the classical method of computing the marginal entropy for all sources. The simulation results are favorable to the proposal.

This work aims to present a new method to perform blind extraction of chaotic deterministic sources mixed with stochastic signals. This technique employs the recurrence quantification analysis (RQA), a tool commonly used to study dynamical systems, to obtain the separating system that recovers the deterministic source. The method is applied to invertible and underdetermined mixture models considering different stochastic sources and different RQA measures. A brief discussion about the influence of recurrence plot parameters on the robustness of the proposal is also provided and illustrated by a set of representative simulations.

In this paper we propose a blind source separation method to process the data acquired by an array of ion-selective electrodes in order to measure the ionic activity of different ions in an aqueous solution. While this problem has already been studied in the past, the method presented differs from the ones previously analyzed by approximating the mixing function by a second-degree polynomial, and using a method based on the differential of the mutual information to adjust the parameter values. Experimental results, both with synthetic and real data, suggest that the algorithm proposed is more accurate than the other models in the literature.

A pragmatic approach for entropy estimation is presented. It is based on coincidence detection and its application leads to a simple algorithm with two attractive features: the method can be used without any a priori knowledge concerning the source alphabet cardinality ( K ) and it can provide useful, though biased, estimates even when the number of samples is less than K .

In this work, we investigate the use of monotonic neural networks as compensating functions in the context of source separation of post-nonlinear PNL mixtures. We first provide a numerical example that illustrates the importance of having bijective nonlinear compensating functions in PNL models. Then, we propose a separation framework in which a monotonic neural network is considered in the first stage of the PNL separating system. Finally, numerical experiments are performed to assess the proposed framework.

Brain–computer interface (BCI) systems based on electroencephalography have been increasingly used in different contexts, engendering applications from entertainment to rehabilitation in a non-invasive framework. In this study, we perform a comparative analysis of different signal processing techniques for each BCI system stage concerning steady state visually evoked potentials (SSVEP), which includes: (1) feature extraction performed by different spectral methods (bank of filters, Welch's method and the magnitude of the short-time Fourier transform); (2) feature selection by means of an incremental wrapper, a filter using Pearson's method and a cluster measure based on the Davies–Bouldin index, in addition to a scenario with no selection strategy; (3) classification schemes using linear discriminant analysis (LDA), support vector machines (SVM) and extreme learning machines (ELM). The combination of such methodologies leads to a representative and helpful comparative overview of robustness and efficiency of classical strategies, in addition to the characterization of a relatively new classification approach (defined by ELM) applied to the BCI-SSVEP systems.

In this work, a measure of similarity based on the matching of multivariate probability density functions (PDFs) is proposed. In consonance with the information theoretic learning (ITL) framework, the affinity comparison between the joint PDFs is performed using a quadratic distance, estimated with the aid of the Parzen window method with Gaussian kernels. The motivation underlying this proposal is to introduce a criterion capable of quantifying, to a significant extent, the statistical dependence present on information sources endowed with temporal and/or spatial structure, like audio, images and coded data. The measure is analyzed and compared with the canonical ITL-based approach — correntropy — for a set of blind equalization scenarios. The comparison includes elements like surface analysis, performance comparison in terms of bit error rate and a qualitative discussion concerning image processing. It is also important to remark that the study includes the application of two computational intelligence paradigms: extreme learning machines and differential evolution. The results indicate that the proposal can be, in some scenarios, a more informative formulation than correntropy.

This work studies the application of non-MSE criteria to adapt the linear readout of Extreme Learning Machines (ELMs) in the context of communication channel equalization. A qualitative and experimental analysis is performed, in terms of bit error rate, optimization surface and decision boundary. The results reached by the ELM-based equalizer, considering three different noise models, did not reveal clear advantages of using criteria based on the concepts of error entropy, correntropy, and the L 1 -norm of the error. Notwithstanding, the observed results motivate a theoretical investigation on the conditions under which the potential discrepancies between the optimal solutions of these criteria may be stressed.

The feature extraction stage is one of the main tasks underlying pattern recognition, and, is particularly important for designing Brain-Computer Interfaces (BCIs), i.e. structures capable of mapping brain signals in commands for external devices. Within one of the most used BCIs paradigms, that based on Steady State Visual Evoked Potentials (SSVEP), such task is classically performed in the spectral domain, albeit it does not necessarily provide the best achievable performance. The aim of this work is to use recurrence-based measures in an attempt to improve the classification performance obtained with a classical spectral approaches for a five-command SSVEP-BCI system. For both recurrence and spectral spaces, features were selected using a cluster measure defined by the Davies-Bouldin index and the classification stage was based on linear discriminant analysis. As the main result, it was found that the threshold \(\varepsilon \) of the recurrence plot, chosen so as to yield a recurrence rate of 2.5 %, defined the key discriminant feature, typically providing a mean classification error of less than 2 % when information from 4 electrodes was used. Such classification performance was significantly better than that attained using spectral features, which strongly indicates that RQA is an efficient feature extraction technique for BCI.

In this work, we present a novel framework for automatic feature selection in brain-computer interfaces (BCIs). The proposal, which manipulates features generated in the frequency domain by an estimate of the power spectral density of the EEG signals, is based on feature optimization (with both binary and real coding) using a state-of-the-art artificial immune network, the cob-aiNet. In order to analyze the performance of the proposed framework, two approaches are adopted: a direct use of the Davies-Bouldin index and the use of metrics associated with the operation of an extreme learning machine (ELM) in the role of a classifier. The results reveal that the proposal has the potential of improving the performance of a BCI system, and also provide elements for an analysis of the spectral content of EEG signals and of the performance of ELMs in motor imagery paradigms.

Echo state networks (ESNs) characterize an attractive alternative to conventional recurrent neural network (RNN) approaches as they offer the possibility of preserving, to a certain extent, the processing capability of a recurrent architecture and, at the same time, of simplifying the training process. However, the original ESN architecture cannot fully explore the potential of the RNN, given that only the second-order statistics of the signals are effectively used. In order to overcome this constraint, distinct proposals promote the use of a nonlinear readout aiming to explore higher-order available information though still maintaining a closed-form solution in the least-squares sense. In this work, we review two proposals of nonlinear readouts - a Volterra filter structure and an extreme learning machine - and analyze the performance of these architectures in the context of two relevant signal processing tasks: supervised channel equalization and chaotic time series prediction. The obtained results reveal that the nonlinear readout can be decisive in the process of aproximating the desired signal. Additionally, we discuss the possibility of combining both ideas of nonlinear readouts and preliminary results indicate that a performance improvement can be attained.

International audience; In this work, we deal with the problem of nonlinear blind source separation (BSS). We propose a new method for BSS in overdetermined linear-quadratic (LQ) mixtures. By exploiting the assumption that the sources are sparse in a transformed domain, we define a framework for canceling the nonlinear part of the mixing process. After that, separation can be conducted by linear BSS algorithms. Experiments with synthetic data are performed to assess the viability of our proposal.

The prediction of seasonal streamflow series is very important in countries where power generation is predominantly done by hydroelectric plants. Echo state networks can be safely regarded as promising tools in forecasting because they are recurrent networks that have a simple and efficient training process based on linear regression. Recently, Boccato et al. proposed a new architecture in which the output layer is built using a principal component analysis and a Volterra filter. This work performs a comparative investigation between the performances of different ESNs in the context of the forecasting of seasonal streamflow series associated with Brazilian hydroelectric plants. Two possible reservoir design approaches were tested with the classical and the Volterra-based output layer structures, and a multilayer perceptron was also included to establish bases for comparison. The obtained results show the relevance of these networks and also contribute to a better understanding of their applicability to forecasting problems.

Extreme Learning Machines (ELMs) and Echo State Networks (ESNs) represent promising alternatives in time series forecasting in view of their intrinsic trade-off between performance and mathematical tractability. Both approaches share a key feature: their supervised parameter adaptation is restricted to the output layer, the remaining synaptic weights being chosen according to a priori unsupervised schemes. This work performs a comparative investigation regarding the performances of a classic ELM and ESNs in the context of the prediction of monthly seasonal streamflow series associated with Brazilian hydroelectric plants. With respect to the ESN, two possible reservoir design approaches are tested, as well as the novel architecture of Boccato et al., which is characterized by the use a Volterra filter and PCA in the readout. Additionally, a MLP is included to establish a base for comparison. Results show the relevance of these architectures in modeling seasonal streamflow series.

Echo state networks (ESNs) can be interpreted as promoting an encouraging compromise between two seemingly conflicting objectives: (i) simplicity of the resulting mathematical model and (ii) capability to express a wide range of nonlinear dynamics. By imposing fixed weights to the recurrent connections, the echo state approach avoids the well-known difficulties faced by recurrent neural network training strategies, but still preserves, to a certain extent, the potential of the underlying structure due to the existence of feedback loops within the dynamical reservoir. Moreover, the overall training process is relatively simple, as it amounts essentially to adapting the readout, which usually corresponds to a linear combiner. However, the linear nature of the output layer may limit the capability of exploring the available information, since higher-order statistics of the signals are not taken into account. In this work, we present a novel architecture for an ESN in which the linear combiner is replaced by a Volterra filter structure. Additionally, the principal component analysis technique is used to reduce the number of effective signals transmitted to the output layer. This idea not only improves the processing capability of the network, but also preserves the simplicity of the training process. The proposed architecture is then analyzed in the context of a set of representative information extraction problems, more specifically supervised and unsupervised channel equalization, and blind separation of convolutive mixtures. The obtained results, when compared to those produced by already proposed ESN versions, highlight the benefits brought by the novel network proposal and characterize it as a promising tool to deal with challenging signal processing tasks.

In this work, we investigate the use of a multimodal search framework to deal with a representative formulation of the Sparse Component Analysis (SCA) problem. The proposed method, which employs an artificial immune network in the role of multimodal optimization tool, is explained and tested in different scenarios. The results are promising and indicate the relevance of using global search tool in SCA, as well as the soundness of the immune-inspired proposal.

In this work, we propose an evolutionary-like approach to the problem of blind adaptive spatial filtering that is based on the decision-directed criterion and on the dopt-aiNet, an artificial immune network conceived to perform multimodal search in dynamic environments. The proposal was tested under static and time-varying undermodeled channel models, and, in all cases, its ability to find and track a solution close to the Wiener global optimum was attested. The obtained results reveal that the dopt-aiNet may decisively enhance the performance of adaptive arrays in scenarios built from elements that are representative of some aspects of real-world communication systems.

Support Vector Clustering (SVC) is a recently proposed clustering methodology with promising performance for high-dimensional and noisy datasets, and for clusters with arbitrary shape. This work addresses the application of SVC, a kernel-based method, in a context in which the channel equalization problem is conceived as a clustering task. The main challenge, in this case, is to perform unsupervised clustering aiming at the design of an optimal Bayesian or a blind prediction-based receiver without resorting to a priori information about the transmission medium. The proposed technique employs a two-stage procedure — a combination between the use of SVC to obtain a first set of clusters and an auxiliary heuristic to help separating eventual multiple clouds contained in a single cluster and attribute centers to them via an iterated local search (ILS) algorithm. The obtained results indicate that kernel methods can be successfully applied to the field of signal processing.

In this work, we address the problem of source separation of post-nonlinear mixtures based on mutual information minimization. There are two main problems related to the training of separating systems in this case: the requirement of entropy estimation and the risk of local convergence. In order to overcome both difficulties, we propose a training paradigm based on entropy estimation through order statistics and on an evolutionary-based learning algorithm. Simulation results indicate the validity of the novel approach.

Modern unorganized machines — extreme learning machines and echo state networks — provide an elegant balance between processing capability and mathematical simplicity, circumventing the difficulties associated with the conventional training approaches of feedforward/recurrent neural networks (FNNs/RNNs). This work performs a detailed investigation of the applicability of unorganized architectures to the problem of seasonal streamflow series forecasting, considering scenarios associated with four Brazilian hydroelectric plants and four distinct prediction horizons. Experimental results indicate the pertinence of these models to the focused task.

In light of the recently proposed generalized correlation function named correntropy, which exploits higher-order statistics and the time structure of signals, we have, in this work, two main objectives: 1) to give a new interpretation --- founded on the relationships between the constant modulus (CM) and Shalvi-Weinstein criteria and between the latter and methods for ICA based on nongaussianity --- to the performance of the constant modulus approach under dependent sources and 2) to analyze the correntropy in the context of blind deconvolution of i.i.d. and dependent sources, as well as to establish elements of a comparison between it and the CMA. The analyses and simulation results unveil some theoretical aspects hitherto unexplored.

In this work, we propose and analyze a method to solve the problem of underdetermined blind source separation (and identification) that employs the ideas of sparse component analysis (SCA) and independent component analysis (ICA). The main rationale of the approach is to allow the possibility of reaching a method that is more robust with respect to the degree of sparseness of the involved signals and more effective in the use of information brought by multiple sensors. The ICA-based solution is tested with the aid of three representative scenarios and its performance is compared with that of one of the soundest SCA techniques available, the DEMIXN algorithm.