Committee Of Neural Networks Research Articles

The Random Hivemind: An ensemble deep learning application to the solar energetic particle prediction problem

The application of machine learning and deep learning techniques, including the wide use of non-ensemble, conventional neural networks (CoNN), for predicting various phenomena has become very popular in recent years thanks to the efficiencies and the abilities of these techniques to find relationships in data without human intervention. However, certain CoNN setups may not work on some datasets, especially if the parameters passed to it, including model parameters and hyperparameters, are arguably arbitrary in nature and need to continuously be updated with the need to retrain the model, especially if the additions of new features render old models obsolete. This concern can be partially alleviated by employing committees of neural networks that are identical in terms of input features and initialized randomly and “vote” on the decisions made by the committees as a whole. Yet, members of the committee have similar architectures and features passed to them, making it possible for the committee members to “agree” on identical sets of weights and biases for all nodes and edges. Members of these committees also cannot be expanded to accommodate new features and entire committees must therefore be retrained in order to do so. We propose the Random Hivemind (RH) approach, which helps to alleviate this concern by having multiple neural network estimators make decisions based on random permutations of features and prescribing a method to determine the weight of the decision of each individual estimator. The effectiveness of RH is demonstrated through experimentation in the predictions of hazardous Solar Energetic Particle (SEP) events by comparing it to that of using both CoNNs and the aforementioned setup of committees identical in input features in this application. Our results demonstrate that RH, while having a comparable or better performance than the CoNN and a Committee-based approach, demonstrates a lesser score spread for the individual experiments, and shows promising results with respect to capturing almost every single flare instance leading to SEPs.

Toward predicting thermal conductivity of hybrid nanofluids: Application of a committee of robust neural networks, theoretical, and empirical models

In this work, three optimized intelligent models on the basis of multilayer perceptron (MLP), two radial basis function (RBF) models (including RBF- particle swarm optimization (PSO) and RBF- farmland fertility algorithm (FFA)), and a committee machine intelligent system (CMIS) were used to estimate Thermal Conductivity (TC) of hybrid nanofluids. Base fluid TC, TC of first and second nanoparticles, size of first and second nanoparticles, volume fractions of first and second nanoparticles, and temperature were considered as inputs to model TC of hybrid nanofluids. Based on the results, the CMIS model emerged as the most accurate model for predicting TC coefficient of hybrid nanofluids, boasting an average absolute percent relative error (AAPRE) of just 1.064%. Comparison with various empirical and theoretical models further highlighted the superior predictive performance of the CMIS model in assessing the TC of hybrid nanofluids. Ultimately, the leverage technique underscored the reliability of the CMIS model.

Strong lensing in UNIONS: Toward a pipeline from discovery to modeling

We present a search for galaxy-scale strong gravitational lenses in the initial 2500 square degrees of the Canada-France Imaging Survey (CFIS). We designed a convolutional neural network (CNN) committee that we applied to a selection of 2 344 002 exquisite-seeingr-band images of color-selected luminous red galaxies. Our classification uses a realistic training set where the lensing galaxies and the lensed sources are both taken from real data, namely the CFISr-band images themselves and theHubbleSpace Telescope (HST). A total of 9460 candidates obtain a score above 0.5 with the CNN committee. After a visual inspection of the candidates, we find a total of 133 lens candidates, of which 104 are completely new. The set of false positives mainly contains ring, spiral, and merger galaxies, and to a lesser extent galaxies with nearby companions. We classify 32 of the lens candidates as secure lenses and 101 as maybe lenses. For the 32 highest quality lenses, we also fit a singular isothermal ellipsoid mass profile with external shear along with an elliptical Sersic profile for the lens and source light. This automated modeling step provides distributions of properties for both sources and lenses that have Einstein radii in the range 0.5″ < θE < 2.5″. Finally, we introduce a new lens and/or source single-band deblending algorithm based on auto-encoder representation of our candidates. This is the first time an end-to-end lens-finding and modeling pipeline is assembled together, in view of future lens searches in a single band, as will be possible withEuclid.

Moment and shear estimations in steel girder highway bridges using committees of artificial neural networks and finite element models

Moment and shear estimations in steel girder highway bridges using committees of artificial neural networks and finite element models

Skyrme-Hartree-Fock-Bogoliubov mass models on a 3D mesh: effect of triaxial shape

The modeling of nuclear reactions and radioactive decays in astrophysical or earth-based conditions requires detailed knowledge of the masses of essentially all nuclei. Microscopic mass models based on nuclear energy density functionals (EDFs) can be descriptive and used to provide this information. The concept of intrinsic symmetry breaking is central to the predictive power of EDF approaches, yet is generally not exploited to the utmost by mass models because of the computational demands of adjusting up to about two dozen parameters to thousands of nuclear masses. We report on a first step to bridge the gap between what is presently feasible for studies of individual nuclei and large-scale models: we present a new Skyrme-EDF-based model that was adjusted using a three-dimensional coordinate-space representation, for the first time allowing for both axial and triaxial deformations during the adjustment process. To compensate for the substantial increase in computational cost brought by the latter, we have employed a committee of multilayer neural networks to model the objective function in parameter space and guide us towards the overall best fit. The resulting mass model BSkG1 is computed with the EDF model independently of the neural network. It yields a root mean square (rms) deviation on the 2457 known masses of 741 keV and an rms deviation on the 884 measured charge radii of 0.024 fm.

Robust affect analysis using committee of deep convolutional neural networks

Robust affect analysis using committee of deep convolutional neural networks

Применение неоднородных сверточных нейронных сетей для построения маршрутов перемещения объектов в пространственных средах с накопленным энергетическим потенциалом

We consider the synthesis of a hybrid neural convolutional network with the modular topology-based architecture, which allows to arrange a parallel convolutional computing system to combine both the energy transfer and data processing, in order to simulate complex functions of natural biological neural populations. The system of interlayer neural commutation, based on the distributed resonance circuits with the layers of electromagnetic metamaterial between the inductive elements, is a base for simulation of the interaction between the astrocyte networks and the neural clusters responsible for information processing. Consequently, the data processing is considered both at the level of signal transmission through neural elements, and as interaction of artificial neurons and astrocytic networks ensuring their functioning. The resulting two-level neural system of data processing implements a set of measures to solve the issue based on the neural network committee. The specific arrangement of the neural network enables us to implement and configure the educational procedure using the properties absent in the neural networks consisting of neural populations only. The training of the convolutional network is based on a preliminary analysis of rhythmic activity, where artificial astrocytes play the main role of interneural switches. The analysis of the signals moving through the neural network enables us to adjust variable components to present information from training bunches in the available memory circuits in the most efficient way. Moreover, in the training process we observe the activity of neurons in various areas to evenly distribute the computational load on neural network modules to achieve maximum performance. The trained and formed convolutional network is used to solve the problem of determining the optimal path for the object moving due to the energy from the environment

Using committees of artificial neural networks with finite element modeling for steel girder bridge load rating estimation

This study examines the suitability of artificial neural networks (ANNs) for refined load rating estimation and bridge management. Bridge management for girder bridges commonly relies on AASHTO line girder (1D) analyses. Less conservative, more rigorous methods are permitted by AASHTO, but provide uncertain return on investment. This study demonstrates that a small set of refined analyses, coupled with ANNs, can be used as a predictive tool to anticipate the likely outcome for similar methods and bridges in a population. Two ANN-based load-rating prediction models were considered: (1) single-best-network, and (2) committee networks (CN). Load rating prediction accuracy was examined on a hybrid subset, consisting of hypothetical and real bridges representative of a steel girder bridge inventory. ANN-based prediction models were trained to map governing inputs (structural and geometric bridge characteristics) to load ratings obtained from 3D FE analyses. Prediction accuracy for bridges outside the training subset demonstrated that refined load ratings can be reliably estimated (about 5% mean absolute error) using a properly trained network model with optimized model complexity. The CN model provided improved prediction accuracy with higher confidence levels than the single-best-network approach, and exhibits less sensitivity to population size as the training sample size was reduced.

Transformer of voice passwords of speakers into a cryptographic key based on a committee of pretrained convolutional neural networks

Transformer of voice passwords of speakers into a cryptographic key based on a committee of pretrained convolutional neural networks

MESING – a new method of organizing the joint work of neural networks and its metrology

A description of a new method for the collective solution of local problems – the method of evolutionary coordination of solutions based on the original use of genetic algorithms – is given. Interaction rules, developed on their basis, coordinate the work of intelligent agents (actors). Based on the Rasch model, an absolute scale for measuring the intellectual power of actors and the costs of intellectual labor when solving local problems with a given probability of their correct solution is introduced. The unit of measurement for these values is introduced and justified – 1 INT. A number of theorems are presented that make it possible to substantiate a new procedure for obtaining a collective solution (mesing) both to increase the intellectual power of a committee of neural networks by 150 times in comparison with a single neural network as part of a committee, and to reduce the probability of an erroneous decision to zero under certain conditions. As a result of the committee’s work, either the correct decision is formed, or the answer “no solution has been found” with a low probability of an erroneous decision.

Application of pool-based active learning in reducing the number of required response history analyses

A step by step method is presented for reducing the need for a large number of response history analyses (RHAs) in developing surrogates to predict the structural responses. These surrogates, which map ground motions features and characteristics of the structural systems into structural responses, are used in deriving fragility curves; and mostly are developed using machine learning algorithms. A machine learning algorithm, depending on the complexity of the model, requires a sufficient amount of training data to predict the outputs accurately. For complicated structural models, generating training data can be computationally demanding. Therefore, there is a need to generate the least amount of training data while preserving the accuracy of the prediction models. Towards this goal, a pool-based query-by-committee active learning (AL) algorithm is applied to choose probably the most informative data samples for surrogate training. A pool of unlabeled GM data samples is generated and a committee of artificial neural networks (ANNs) is defined to choose the smallest subset of the data that would be the most informative training data. The selected data samples meet informativeness, representativeness, and diversity criteria. The results of the applied case study show that implementing AL, can considerably reduce the size of required training data and consequently the amount of RHAs while improving the performance of the prediction models. Specifically, the findings demonstrate that AL improves the average F-scores of ANNs by about 10% and significantly reduces their variations, which indicates its stable behavior.

Taming Nuclear Complexity with a Committee of Multilayer Neural Networks.

We demonstrate that a committee of deep neural networks is capable of predicting the ground-state and excited energies of more than 1800 atomic nuclei with an accuracy akin to the one achieved by state-of-the-art nuclear energy density functionals (EDFs) and with significantly less computational cost. An active learning strategy is proposed to train this algorithm with a minimal set of 210 nuclei. This approach enables future fast studies of the influence of EDF parametrizations on structure properties over the whole nuclear chart and suggests that for the first time a machine learning framework successfully encoded several correlated aspects of nuclear deformation.

DEVELOPMENT OF EXCITATION STRUCTURE RBF-METAMODELS OF MOVING CONCENTRIC EDDY CURRENT PROBE

Introduction. The work is devoted to metamodels creation of surface circular concentric eddy current probe. Formulation of the problem. In the problem of surface circular concentric eddy current probe synthesis in the general formulation, apriori given desired eddy currents density distribution in the control zone was used. The realization of the optimal synthesis problem involves a multiple solution to the analysis problem for each current structure of numerical calculations excitation, which are very costly in terms of computational and time costs, which makes it impossible to solve the synthesis problem in the classical formulation. By solving the critical resource intensiveness problem, there is the surrogate optimization technology using of that uses the surface circular concentric eddy current probe metamodel, which is much simpler in realization and is an approximation of the exact electrodynamic model. Goal. Creation of surface circular concentric eddy current probe RBF-metamodels, which can be used to calculate eddy currents density distribution in the control zone and suitable for use in optimal synthesis problems. Method. To develop an approximation model, a mathematical apparatus for artificial neural networks, namely, RBF–networks, has been used, whose accuracy has been increased with the help of the neural networks committee. Correction of errors in the committee was reduced by applying the bagging procedure. During the network training the regularization technique is used, which avoids re-learning the neural network. The computer experiment plan was performed using the Sobol LPt–sequences. The obtained multivariable regression model quality evaluation was performed by checking the response surface reproducibility correctness in the entire region of variables variation. Results. The modelling of eddy currents density distribution calculations on exact electrodynamic mathematical models in the experimental plan points are carried out. For the immovable and moving surface circular concentric eddy current probe, RBF–metamodels were constructed with varying spatial coordinates and radius. Scientific novelty. Software was developed for eddy currents density distribution calculation in the surface circular concentric eddy current probe control zone taking into account the speed effect on exact electrodynamic mathematical models and for forming experiment plan points using the Sobol LPt–sequences. The geometric surface circular concentric eddy current probe excitation structures models with homogeneous sensitivity for their optimal synthesis taking into account the speed effect are proposed. Improved computing technology for constructing metamodels. The RBF-metamodels of the surface circular concentric eddy current probe are built and based on the speed effect. Practical significance. The work results can be used in the surface circular concentric eddy current probe synthesis with an apriori given eddy currents density distribution in the control zone.

Classifcation of events in information security systems based on neural networks

Purpose of the research.The aim of the study is to increase the effectiveness of information security and to enhance accuracy and promptness of the classification of security events, security incidents, and threats in information security systems. To respond to this challenge, neural network technologies were suggested as a classification tool for information security systems. These technologies allow accommodating incomplete, inaccurate and unidentified raw data, as well as utilizing previously accumulated information on security issues. To address the problem more effectively, collective methods based on collective neural ensembles aligned with an advanced complex approach were implemented.Materials and methods:When solving complex classification problems, often none of the classification algorithms provides the required accuracy. In such cases, it seems reasonable to build compositions of algorithms, mutually compensating errors of individual algorithms. The study also gives an insight into the application of neural network ensemble to address security issues in the corporate information system and provides a brief review of existing approaches to the construction of neural network ensembles and methods to shape problem solving with neural networks classifiers. An advanced integrated approach is proposed to tackle problems of security event classification based on neural network ensembles (neural network committees). The approach is based on a three-step procedure. The stages of the procedure implementation are described. It is shown that the use of this approach facilitates the efficiency of solving the problem.Results:An advanced integrated approach to addressing security event classification based on neural network ensembles (neural network committees) is proposed. This approach applies adaptive reduction of neural network ensemble (selection of the best classifiers is based on the assessment of the compliance degree of the competence area of the private neural network classifier and convergence of the results of private classifiers), as well as the selection and rationale of the voting method (composition or aggregation of outputs of private classifiers). The results of numerical experiments support the effectiveness of the proposed approach.Conclusion:Collectively used artificial neural networks in the form of neural network ensembles (committees of neural networks) will provide more accurate and reliable results of security event classification in the corporate information network. Moreover, an advanced integrated approach to the construction of a neural network ensemble is proposed to facilitate effectiveness of the classification process. The approach is based on the application of the adaptive reduction procedure for the results of private classifiers and the procedure for selecting the method of aggregation of the results of private classifiers. These outcomes will enable advancement of the system control over information security incidents. Finally, the paper defines tendencies and directions of the development of collective solution methods applying neural network ensembles (committees of neural networks).

Hybridizing Evolutionary Computation and Deep Neural Networks: An Approach to Handwriting Recognition Using Committees and Transfer Learning

Neuroevolution is the field of study that uses evolutionary computation in order to optimize certain aspect of the design of neural networks, most often its topology and hyperparameters. The field was introduced in the late‐1980s, but only in the latest years the field has become mature enough to enable the optimization of deep learning models, such as convolutional neural networks. In this paper, we rely on previous work to apply neuroevolution in order to optimize the topology of deep neural networks that can be used to solve the problem of handwritten character recognition. Moreover, we take advantage of the fact that evolutionary algorithms optimize a population of candidate solutions, by combining a set of the best evolved models resulting in a committee of convolutional neural networks. This process is enhanced by using specific mechanisms to preserve the diversity of the population. Additionally, in this paper, we address one of the disadvantages of neuroevolution: the process is very expensive in terms of computational time. To lessen this issue, we explore the performance of topology transfer learning: whether the best topology obtained using neuroevolution for a certain domain can be successfully applied to a different domain. By doing so, the expensive process of neuroevolution can be reused to tackle different problems, turning it into a more appealing approach for optimizing the design of neural networks topologies. After evaluating our proposal, results show that both the use of neuroevolved committees and the application of topology transfer learning are successful: committees of convolutional neural networks are able to improve classification results when compared to single models, and topologies learned for one problem can be reused for a different problem and data with a good performance. Additionally, both approaches can be combined by building committees of transferred topologies, and this combination attains results that combine the best of both approaches.

Design and field implementation of an impact detection system using committees of neural networks

Many critical societal functions depend on uninterrupted service of civil engineering infrastructure. Railroads represent important infrastructure components of the transportation sector and provide both passenger and freight services. Railroad bridges over roadways are susceptible to impacts from overheight vehicles and equipment, which may damage bridge girders or supports and must be investigated after each event. One method of monitoring for vehicle-bridge collisions utilizes accelerometers to monitor for abnormal bridge vibrations corresponding to abnormal activity. Passing trains under normal operating conditions frequently produce significant bridge responses that have similar response characteristics to bridge strikes, but do not need to be investigated. This paper presents an expert system which comprises committees of artificial neural networks trained to interrogate data collected from accelerometers mounted on the bridge, assess the nature of the acceleration signal, and classify the event as either a passing train or a potentially damaging impact. This system is trained using acceleration time histories from accelerometers installed on 8 low-clearance rail bridges; no finite element model simulations were used for network training or data stream creation. The presented system accurately detects and classifies impacts with average impact detection performance ranging from 91–100% with average false positive rates limited to 0.00–0.75%.

Model Updating and Crack Detection for Elastically Restrained Tapered Cantilever-Type Beam Using Natural Frequencies

The current paper presents a method to update the baseline model and detect cracks in tapered cantilever pipe-type beams which are restrained by a translational and rotational spring with a tip mass at the free end using natural frequencies. Modal parameters of intact beam are obtained by applying boundary conditions to a general solution for tapered beam. The equivalent bending stiffness is used to calculate natural frequencies of cracked beams. An experimental study is carried out. The translational and rotational spring constants are updated to establish a baseline model using a neural network. Then, several numerical analyses of the cracked beams are carried out to extract the natural frequencies, and those are used in constructing the training patterns of a neural network. The committee of neural networks are employed to identify the cracks. The crack identifications are carried out for the 3 damage cases, and it is found that the estimated crack locations and sizes agree reasonably well with the exact values.

Four-dimensional Spherical Model of Emotions

The four-dimensional spherical model of emotions, which is anthropomorphic and neurotropic because it reflects a subjective attitude of the person as well as the work of the neural systems, is presented as a universal "measure" in organizing the committee of neural networks. The article shows the necessity of standardizing the outputs of individual local analyzers (neural networks, forming the committee) according to the dimensionality and "code" as it is realized in the motor system of vertebrates. The proposed four-dimensional model can be used as a general classification system for emotional phenomena, combining both physiological concepts of the brain mechanisms of emotional regulation and known psychological classifications, obtained on the basis of different experimental data. As a result, this model is anthropomorphic (because it reflects the subjective attitude of a person) and neurotropic (because it reflects neural mechanisms) and it allows to quantitatively describe and visually present changes in the current state of a person or their emotional relationship. The model can become a basis in constructing a device, based on biologically inspired cognitive architectures, to become a "chairperson" in the neural network committee.

Experiments using deep learning for dermoscopy image analysis

Abstract Skin cancer is a major public health problem, as is the most common type of cancer and represents more than half of cancer diagnosed worldwide. Early detection influences the outcome of the disease and motivates the research presented in this paper. Recent results show that deep learning based approaches learn from data, and can outperform human specialists in a set of tasks when large databases are available for training. This research investigates the scenario where the amount of data available for training is small. It obtains relevant results for the ISBI 2016 melanoma classification challenge (named Skin Lesion Analysis for Melanoma Detection) facing the peculiarities of dealing with such a small and unbalanced biological database. To do this, it explores committees of Deep Convolutional Neural Networks (DCNN), the augmentation of the training data set by image processing classical transforms and by deformations guided by expert knowledge about the lesion axis, and it introduces a third class aiming to improve the classifiers’ distinction of the region of interest of the lesion. The experiments show that the proposed approach improves the final classifier invariance for common melanoma variations, common skin patterns and markers, and dermatoscope capturing conditions.

A particle filter-based model selection algorithm for fatigue damage identification on aeronautical structures

The early diagnosis of cracks in aeronautical structures is a fundamental task for the safe system operation and the optimization of maintenance policies, in view of the increasing interest in life extension programs of several high-investment industries. In principle, these tasks could be fulfilled within a condition-based framework, where direct or indirect observations of the degradation evolution are processed, possibly in real time, by proper diagnostic computational tools. In the past, several attempts have been made to build real-time monitoring systems collecting strain signals acquired from sensor networks. However, in real applications, some issues remain unresolved, for example, the large number of observations available to be handled within a unique diagnostic framework, their relationship with the underlying crack size, and their typical large uncertainties. In this paper, we provide a practical solution by innovatively combining a particle filtering-based model identification algorithm with a measurement system exploiting real-time observations of the crack length reconstructed by a committee of artificial neural networks. The artificial neural networks are trained by simulated strain fields generated by a finite element model. The resulting tool allows to perform an automatic, simultaneous, and real-time (a) selection of the model more properly describing the system state evolution, so as to detect the crack propagation onset time, (b) estimation of the model parameters, and (c) estimation of the crack length, within a unique probabilistic framework based on particle filtering. The methodology is demonstrated with reference to a real helicopter panel subject to fatigue and equipped with a fiber Bragg grating sensor network.