Electroencephalography (EEG) datasets are often small and high dimensional, owing to cumbersome recording processes. In these conditions, powerful machine learning techniques are essential to deal with the large amount of information and overcome the curse of dimensionality. Artificial Neural Networks (ANNs) have achieved promising performance in EEG-based Brain-Computer Interface (BCI) applications, but they involve computationally intensive training algorithms and hyperparameter optimization methods. Thus, an awareness of the quality-cost trade-off, although usually overlooked, is highly beneficial. In this paper, we apply a hyperparameter optimization procedure based on Genetic Algorithms to Convolutional Neural Networks (CNNs), Feed-Forward Neural Networks (FFNNs), and Recurrent Neural Networks (RNNs), all of them purposely shallow. We compare their relative quality and energy-time cost, but we also analyze the variability in the structural complexity of networks of the same type with similar accuracies. The experimental results show that the optimization procedure improves accuracy in all models, and that CNN models with only one hidden convolutional layer can equal or slightly outperform a 6-layer Deep Belief Network. FFNN and RNN were not able to reach the same quality, although the cost was significantly lower. The results also highlight the fact that size within the same type of network is not necessarily correlated with accuracy, as smaller models can and do match, or even surpass, bigger ones in performance. In this regard, overfitting is likely a contributing factor since deep learning approaches struggle with limited training examples.

This paper presents a new wrapper method able to optimize simultaneously the parameters of the classifier while the size of the subset of features that better describe the input dataset is also being minimized. The search algorithm used for this purpose is based on a co-evolutionary algorithm optimizing several objectives related with different desirable properties for the final solutions, such as its accuracy, its final number of features, and the generalization ability of the classifier. Since these objectives can be sorted according to their priorities, a lexicographic approach has been applied to handle this many-objective problem, which allows the use of a simple evolutionary algorithm to evolve each one of the different sub-populations.

Feature selection is an important step in building classifiers for high-dimensional data problems, such as EEG classification for BCI applications. This paper proposes a new wrapper method for feature selection, based on a multi-objective evolutionary algorithm, where the representation of the individuals or potential solutions, along with the breeding operators and objective functions, have been carefully designed to select a small subset of features that has good generalization capability, trying to avoid the over-fitting problems that wrapper methods usually suffer. A novel feature ranking procedure is also proposed in order to analyze the stability of the proposed wrapper method. Four different classification schemes have been applied within the proposed wrapper method in order to evaluate its accuracy and stability for feature selection on a real motor imagery dataset. Experimental results show that the wrapper method presented in this paper is able to obtain very small subsets of features, which are quite stable and also achieve high classification accuracy, regardless of the classifiers used. Crown Copyright (C) 2019 Published by Elsevier B.V. All rights reserved.

This paper proposes a supervised filter method for evolutionary multi-objective feature selection for classification problems in high-dimensional feature space, which is evaluated by comparison with wrapper approaches for the same application. The filter method based on a set of label-aided utility functions is compared with wrapper approaches using the accuracy and generalization properties in the effective searching of the most adequate subset of features through an evolutionary multi-objective optimization scheme. The target application corresponds to a brain–computer interface (BCI) classification task based on linear discriminant analysis (LDA) classifiers, where the properties of multi-resolution analysis (MRA) for signal analysis in temporal and spectral domains have been used to extract features from electroencephalogram (EEG) signals. The results, corresponding to a dataset obtained from the databases of the BCI Laboratory of the University of Essex, UK, including ten subjects with three different imagery movements, have allowed us to evaluate the advantages and drawbacks of the different approaches with respect to time consumption, accuracy and generalization capabilities.

High-dimensional pattern classification problems with a small number of training patterns are difficult. This paper deals with classification of motor imagery tasks for brain-computer interfacing (BCI), which is a hard problem involving a relatively small number of high-dimensional training patterns where curse of dimensionality issue has to be taken into account and feature selection is an important requirement to build a suitable classifier. Evolutionary metaheuristics for feature selection are usually more time-consuming than other alternatives, but their high performances in terms of classification accuracy make them desirable approaches. In this paper, feature selection through a wrapper procedure based on multi-objective optimization is compared with the use of deep belief networks (DBN) that constitute powerful classifiers implementing feature selection implicitly. Two different classifiers, LDA (linear discriminant analysis) and DBN, have been used to classify EEG signals with features extracted by multiresolution analysis (MRA) and selected by a multiobjective evolutionary method that also uses LDA to implement the fitness function of the solutions. The experimental results show that DBNs usually provide better or similar classification performances without requiring an explicit feature selection phase. Nevertheless, the DBN’s classification performance significantly decreases in problems with a very large number of features. Moreover, to achieve high classification rates, it is necessary to determine a suitable structure for the DBN. Therefore, in this paper we also propose a multiobjective approach to tackle this problem.

The reduction of energy consumption in buildings is one of the goals to improve energy efficiency. One way to achieve energy savings in buildings is to develop intelligent control strategies for heating systems that are able to reduce power consumption without affecting the thermal comfort. An intelligent control system must be able to predict the temperature of the building in order to manage the heating system. In this paper, we present a rule-based model that is able to predict the indoor temperature for different values of k (hours ahead in time). The model has been learned with FRULER, a genetic fuzzy system that generates accurate and simple knowledge bases. Our approach has been validated with real data from a residential college.

This paper proposes and evaluates a filter approach for evolutionary multi-objective feature selection in classification problems with a large number of features. Such classification problems frequently appear in many bioinformatics applications where the number of patterns is smaller than the number of features and thus the curse of dimensionality problem exists. The main contribution of this paper is proposing a set of label-aided utility functions that allows the effective search of the most adequate subset of features through an evolutionary multi-objective optimization scheme. The experimental results have been obtained in a brain-computer interface (BCI) classification task based on LDA classifiers, where the properties of multi-resolution analysis (MRA) for signal analysis in temporal and spectral domains have been used to extract the features from EEG signals. The results from the proposed filter method demonstrate some advantages such as less time consumption and better generalization capabilities with respect to some wrapper-based multi-objective feature selection alternatives.

In the context of supervised vectorial quantization (VQ) learning algorithms, we present an algorithm (SLTI) that exploits the self-organizing properties arising from a particular process of temporal inhibition of the winning units in competitive learning. This exploitation consists of establishing independence capabilities in the initialization of the prototypes (weight vectors), together with generalization capabilities, which to a certain extent solve some of the critical problems involved in the use of conventional algorithms such as LVQs and DSM. Another original aspect of this paper is the inclusion in SLTI of a simple rule for prototype adaptation, which incorporates certain useful features that make possible to plan the configuration of the SLTI parameters with specific goals in order to approach classification tasks of varied complexity and natures (versatility). This versatility is experimentally demonstrated with synthetic data comprising non linearly-separable classes, overlapping classes and interlaced classes with a certain degree of overlap.

Biologically inspired VLSI circuits are proposed which can be particularized to approximate the real-time dynamics of either additive or shunting neural models. Analog inputs to these circuits are represented by short spikes and, both, their transient and steady-state behaviours depend only on process-independent local ratios. The paper includes simulation results and experimental measures of a CMOS prototype, which illustrate the utility and the feasibility of the proposed VLSI approach.

In this paper, a learning algorithm that leads to an efficient self-organization in a Kohonen Neural Network (KNN) with fixed neighbourhood is presented. This algorithm may be faster than the originally proposed for KNNs, produces in general better covering of the input stimulus space, and can be more easily implemented in hardware due to the fixed neighbourhood it manages.

This communication presents an Extended Hopfield Neural Network which has been applied to design the extra circuitry for testing a digital circuit during its normal operation, problem which we have shown to be equivalent to the problem of selecting an optimal set of Reed-Muller spectral coefficients. It has been suggested that neural networks, in particular the Hopfield Neural network, may be used to solve linear programming problems. Here, we show how a modification of the Hopfield Network structure also allows to solve non-linear programming problems.

An analog VLSI implementation of a neural network is presented which has been designed for use in clustered systems with competitive learning. The circuit implements an inhibitory cluster that includes the winner-unit computation. The synaptic weights are externally alterable asynchronously with network operation. A test chip has been designed with the rules of a 2- mu m CMOS process which shows high integration density (about 200 synaptic connections per square millimeter). Simulation results and VLSI realization details of different modules comprised in the chip are also presented.<>

Presents a new kind of hierarchical binary network which may be considered as a specific case of cellular networks. Such networks are defined by primitives or processes which are personalised for network implementation in specific applications. The way in which a network may be described is provided by a process (network process) which may be used both for simulation software and for deducing the general hardware architecture of the network. Following a description of the algorithms of different simulation primitives, a network example for noise elimination in images is included to illustrate the potential of this kind of network

A model is presented for a neural network with competitive learning that demonstrates the self-organizing capabilities arising from the inclusion of a simple temporal inhibition mechanism within the neural units. This mechanism consists of the inhibition, for a certain time, of the neuron that generates an action potential; such a process is termed Post_Fire inhibition. The neural inhibition period, or degree of inhibition, and the way it is varied during the learning process, represents a decisive factor in the behaviour of the network, in addition to constituting the main basis for the exploitation of the model. Specifically, we show how Post_Fire inhibition is a simple mechanism that promotes the participation of and cooperation between the units comprising the network; it produces self-organized neural responses that reveal spatio–temporal characteristics of input data. Analysis of the inherent properties of the Post_Fire inhibition and the examples presented show its potential for applications such as vector quantization, clustering, pattern recognition, feature extraction and object segmentation. Finally, it should be noted that the Post_Fire inhibition mechanism is treated here as an efficient abstraction of biologically plausible mechanisms, which simplifies its implementation.

A new method for separating linear mixtures of statistically independent signals with super-Gaussian probability distributions, using a simple neural network, is proposed. The procedure is based on geometric properties, and it is shown that the maxima of the mixed density distribution belong to straight lines, the direction vectors of which, when taken as columns of a matrix, comprise a demixing matrix. The results obtained with synthetic mixtures of real speech signals are shown.

The paper describes the implementation of competitive neural structures based on a spiking neural model that includes multiplicative or shunting synapses enabling non-saturated stable states in response to different stationary inputs as well as controllable transient responses. A VLSI-viable implementation of this model has been previously proposed and tested [1]. It has the possibility of modulating the output spike frequency by an additional input without affecting other neuron variables such as the membrane potential. This feature is exploited in the simulation of a Selective Temporal Inhibition network that is suitable for implementing attentional control systems.

A cost-effective audit system has been developed that will both detect systematic error in data and procedures and evaluate the quality assurance programme provided by a physics department for radiotherapy. The audit has been developed for external beam radiotherapy and assesses one modality and one treatment machine per year. The audit is carried out on an interdepartmental basis and can be undertaken by two physicists from each department in one working day. The method of assessing the quality assurance programme and the schedule of measurements are described. The process is illustrated using the results of trial audits between the medical physics departments at Coventry and Leicester.

A synthesis procedure is described for multivalued threshold decoders and their application to the implementation of multivalued functions. This method is based on the definition of inversion and extreme functions. Decoders are obtained using only the very simple blocks which synthesize inversion functions. This general design procedure may be used for any integrated circuit technology. As an example, the method is applied to the synthesis of an integrated CMOS quaternary decoder. Both the layout and the timing of the integrated circuit, and their application to the synthesis of multivalued functions and multistable memory elements, are shown.