• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.12
• /
• pp.5765-5781
• /
• 2018

Extreme learning machine (ELM) is emerging as a powerful machine learning method in a variety of application scenarios due to its promising advantages of high accuracy, fast learning speed and easy of implementation. However, how to select the optimal hidden layer of ELM is still an open question in the ELM community. Basically, the number of hidden layer nodes is a sensitive hyperparameter that significantly affects the performance of ELM. To address this challenging problem, we propose to adopt multiple kernel learning (MKL) to design a multi-hidden-layer-kernel ELM (MHLK-ELM). Specifically, we first integrate kernel functions with random feature mapping of ELM to design a hidden-layer-kernel ELM (HLK-ELM), which serves as the base of MHLK-ELM. Then, we utilize the MKL method to propose two versions of MHLK-ELMs, called sparse and non-sparse MHLK-ELMs. Both two types of MHLK-ELMs can effectively find out the optimal linear combination of multiple HLK-ELMs for different classification and regression problems. Experimental results on seven data sets, among which three data sets are relevant to classification and four ones are relevant to regression, demonstrate that the proposed MHLK-ELM achieves superior performance compared with conventional ELM and basic HLK-ELM.

• ETRI Journal
• /
• v.15 no.2
• /
• pp.35-51
• /
• 1993

This paper presents function approximation based on nonparametric estimation. As an estimation model of function approximation, a three layered network composed of input, hidden and output layers is considered. The input and output layers have linear activation units while the hidden layer has nonlinear activation units or kernel functions which have the characteristics of bounds and locality. Using this type of network, a many-to-one function is synthesized over the domain of the input space by a number of kernel functions. In this network, we have to estimate the necessary number of kernel functions as well as the parameters associated with kernel functions. For this purpose, a new method of parameter estimation in which linear learning rule is applied between hidden and output layers while nonlinear (piecewise-linear) learning rule is applied between input and hidden layers, is considered. The linear learning rule updates the output weights between hidden and output layers based on the Linear Minimization of Mean Square Error (LMMSE) sense in the space of kernel functions while the nonlinear learning rule updates the parameters of kernel functions based on the gradient of the actual output of network with respect to the parameters (especially, the shape) of kernel functions. This approach of parameter adaptation provides near optimal values of the parameters associated with kernel functions in the sense of minimizing mean square error. As a result, the suggested nonparametric estimation provides an efficient way of function approximation from the view point of the number of kernel functions as well as learning speed.

• Journal of Electrical Engineering and Technology
• /
• v.11 no.4
• /
• pp.993-1002
• /
• 2016

Hepatitis is a major public health problem all around the world. This paper proposes an automatic disease diagnosis system for hepatitis based on Genetic Algorithm (GA) Wavelet Kernel (WK) Extreme Learning Machines (ELM). The classifier used in this paper is single layer neural network (SLNN) and it is trained by ELM learning method. The hepatitis disease datasets are obtained from UCI machine learning database. In Wavelet Kernel Extreme Learning Machine (WK-ELM) structure, there are three adjustable parameters of wavelet kernel. These parameters and the numbers of hidden neurons play a major role in the performance of ELM. Therefore, values of these parameters and numbers of hidden neurons should be tuned carefully based on the solved problem. In this study, the optimum values of these parameters and the numbers of hidden neurons of ELM were obtained by using Genetic Algorithm (GA). The performance of proposed GA-WK-ELM method is evaluated using statical methods such as classification accuracy, sensitivity and specivity analysis and ROC curves. The results of the proposed GA-WK-ELM method are compared with the results of the previous hepatitis disease studies using same database as well as different database. When previous studies are investigated, it is clearly seen that the high classification accuracies have been obtained in case of reducing the feature vector to low dimension. However, proposed GA-WK-ELM method gives satisfactory results without reducing the feature vector. The calculated highest classification accuracy of proposed GA-WK-ELM method is found as 96.642 %.

• ETRI Journal
• /
• v.39 no.6
• /
• pp.841-850
• /
• 2017

We investigate neural network image reconstruction for magnetic particle imaging. The network performance strongly depends on the convolution effects of the spectrum input data. The larger convolution effect appearing at a relatively smaller nanoparticle size obstructs the network training. The trained single-layer network reveals the weighting matrix consisting of a basis vector in the form of Chebyshev polynomials of the second kind. The weighting matrix corresponds to an inverse system matrix, where an incoherency of basis vectors due to low convolution effects, as well as a nonlinear activation function, plays a key role in retrieving the matrix elements. Test images are well reconstructed through trained networks having an inverse kernel matrix. We also confirm that a multi-layer network with one hidden layer improves the performance. Based on the results, a neural network architecture overcoming the low incoherence of the inverse kernel through the classification property is expected to become a better tool for image reconstruction.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.25 no.1
• /
• pp.97-104
• /
• 2015

An 'associative cube', a class of auto-associative memories, is revisited here, in which training data and hidden orthogonal basis functions such as wavelet packets or Fourier kernels, are combined in the weight cube. This weight cube has hidden units in its depth, represented by a three dimensional cubic structure. We develop an unsupervised incremental learning mechanism based upon the adaptive least squares method. Training data are mapped into orthogonal basis vectors in a least-squares sense by updating the weights which minimize an energy function. Therefore, a prescribed orthogonal kernel is incrementally assigned to an incoming data. Next, we show how a decoding procedure finds the closest one with a competitive network in the hidden layer. As noisy test data are applied to an associative cube, the nearest one among the original training data are restored in an optimal sense. The simulation results confirm robustness of associative cubes even if test data are heavily distorted by various types of noise.

• Journal of the Institute of Convergence Signal Processing
• /
• v.7 no.2
• /
• pp.53-59
• /
• 2006

Dynamic neural networks have been applied to diverse fields requiring temporal signal processing such as system identification and signal prediction. This paper proposes the gamma neural network(GAM), which uses gamma memory kernel in the hidden layer of feedforward multilayered network, to improve dynamics of networks and then describes nonlinear adaptive prediction using the proposed network as an adaptive filter. The proposed network is evaluated in nonlinear signal prediction and compared with feedforword(FNN) and recurrent neural networks(RNN) for the relative comparison of prediction performance. Simulation results show that the GAM network performs better with respect to the convergence speed and prediction accuracy, indicating that it can be a more effective prediction model than conventional multilayered networks in nonlinear prediction for nonstationary signals.

• Journal of the Semiconductor & Display Technology
• /
• v.22 no.3
• /
• pp.142-148
• /
• 2023

In this paper, we identify performance issues in executing compute kernels from PolyBench, which includes compute kernels that are the core computational units of various data-intensive workloads, such as deep learning and data-intensive applications, on Processing-in-Memory (PIM) devices. Therefore, using our in-house simulator, we measured and compared the various performance metrics of workloads based on traditional out-of-order and in-order processors with Processing-in-Memory-based systems. As a result, the PIM-based system improves performance compared to other computing models due to the short-term data reuse characteristic of computational kernels from PolyBench. However, some kernels perform poorly in PIM-based systems without a multi-layer cache hierarchy due to some kernel's long-term data reuse characteristics. Hence, our evaluation and analysis results suggest that further research should consider dynamic and workload pattern adaptive approaches to overcome performance degradation from computational kernels with long-term data reuse characteristics and hidden data locality.

• Journal of Biomedical Engineering Research
• /
• v.19 no.5
• /
• pp.469-476
• /
• 1998

In the case of acquisition and displaying medical Images such as ultrasonography and endoscopy on VGA monitor of PC system, image degradation of tear-drop appears through scan conversion. In this study, we compare several methods which can solve this degradation and implement the hardware system that resolves this problem in real-time with PC. It is possible to represent high quality image display and real-time processing and acquisition with specific de-interlacing device and PCI bridge on our hardware system. Image quality is improved remarkably on our hardware system. It is implemented as PC-based system, so acquiring, saving images and describing text comment on those images and PACS networking can be easily implemented.metabolism. All images were spatially normalized to MNI standard PET template and smoothed with 16mm FWHM Gaussian kernel using SPM96. Mean count in cerebral region was normalized. The VOls for 34 cerebral regions were previously defined on the standard template and 17 different counts of mirrored regions to hemispheric midline were extracted from spatially normalized images. A three-layer feed-forward error back-propagation neural network classifier with 7 input nodes and 3 output nodes was used. The network was trained to interpret metabolic patterns and produce identical diagnoses with those of expert viewers. The performance of the neural network was optimized by testing with 5~40 nodes in hidden layer. Randomly selected 40 images from each group were used to train the network and the remainders were used to test the learned network. The optimized neural network gave a maximum agreement rate of 80.3% with expert viewers. It used 20 hidden nodes and was trained for 1508 epochs. Also, neural network gave agreement rates of 75~80% with 10 or 30 nodes in hidden layer. We conclude that artificial neural network performed as well as human experts and could be potentially useful as clinical decision support tool for the localization of epileptogenic zones.

• Journal of the Institute of Convergence Signal Processing
• /
• v.9 no.3
• /
• pp.238-244
• /
• 2008

Dynamic neural networks have been applied to diverse fields requiring temporal signal processing. This paper presents gamma neural network(GAM) to improve the dynamics of multilayer network. The GAM network uses the gamma memory kernel in the hidden layer of feedforword multilayer network. The GAM network is evaluated in linear and nonlinear system identification, and compared with feedforword(FNN) and recurrent neural networks(RNN) for the relative comparison of its performance. Experimental results show that the GAM network performs better with respect to the convergence and accuracy, indicating that it can be a more effective network than conventional multilayer networks in system identification.

• Journal of Biomedical Engineering Research
• /
• v.19 no.5
• /
• pp.455-468
• /
• 1998

For the objective interpretation of cerebral metabolic patterns in epilepsy patients, we developed computer-aided classifier using artificial neural network. We studied interictal brain FDG PET scans of 257 epilepsy patients who were diagnosed as normal(n=64), L TLE (n=112), or R TLE (n=81) by visual interpretation. Automatically segmented volume of interest (VOI) was used to reliably extract the features representing patterns of cerebral metabolism. All images were spatially normalized to MNI standard PET template and smoothed with 16mm FWHM Gaussian kernel using SPM96. Mean count in cerebral region was normalized. The VOls for 34 cerebral regions were previously defined on the standard template and 17 different counts of mirrored regions to hemispheric midline were extracted from spatially normalized images. A three-layer feed-forward error back-propagation neural network classifier with 7 input nodes and 3 output nodes was used. The network was trained to interpret metabolic patterns and produce identical diagnoses with those of expert viewers. The performance of the neural network was optimized by testing with 5~40 nodes in hidden layer. Randomly selected 40 images from each group were used to train the network and the remainders were used to test the learned network. The optimized neural network gave a maximum agreement rate of 80.3% with expert viewers. It used 20 hidden nodes and was trained for 1508 epochs. Also, neural network gave agreement rates of 75~80% with 10 or 30 nodes in hidden layer. We conclude that artificial neural network performed as well as human experts and could be potentially useful as clinical decision support tool for the localization of epileptogenic zones.