Search results for: Neural networks and Fractals.

Authors: Rohitash Chandra, Christian W. Omlin

Abstract:

We present a hybrid architecture of recurrent neural networks (RNNs) inspired by hidden Markov models (HMMs). We train the hybrid architecture using genetic algorithms to learn and represent dynamical systems. We train the hybrid architecture on a set of deterministic finite-state automata strings and observe the generalization performance of the hybrid architecture when presented with a new set of strings which were not present in the training data set. In this way, we show that the hybrid system of HMM and RNN can learn and represent deterministic finite-state automata. We ran experiments with different sets of population sizes in the genetic algorithm; we also ran experiments to find out which weight initializations were best for training the hybrid architecture. The results show that the hybrid architecture of recurrent neural networks inspired by hidden Markov models can train and represent dynamical systems. The best training and generalization performance is achieved when the hybrid architecture is initialized with random real weight values of range -15 to 15.

Keywords: Deterministic finite-state automata, genetic algorithm, hidden Markov models, hybrid systems and recurrent neural networks.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1845

Authors: N.Vahdatifar, Ss.Mortazavi, R.Kianinezhad

Abstract:

In this paper, a Neural Network based predictive DTC algorithm is proposed .This approach is used as an alternative to classical approaches .An appropriate riate Feed - forward network is chosen and based on its value of derivative electromagnetic torque ; optimal stator voltage vector is determined to be applied to the induction motor (by inverter). Moreover, an appropriate torque and flux observer is proposed.

Keywords: Neural Networks, Predictive DTC

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1345

Authors: Christian Mayr, Matthias Ehrlich, Stephan Henker, Karsten Wendt, René Schüffny

Abstract:

Traditionally, VLSI implementations of spiking neural nets have featured large neuron counts for fixed computations or small exploratory, configurable nets. This paper presents the system architecture of a large configurable neural net system employing a dedicated mapping algorithm for projecting the targeted biology-analog nets and dynamics onto the hardware with its attendant constraints.

Keywords: Large scale VLSI neural net, topology mapping, complex pulse communication.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1638

Authors: Samy Sadek, Ayoub Al-Hamadi, Bernd Michaelis, Usama Sayed

Abstract:

In this paper, we propose a supervised method for color image classification based on a multilevel sigmoidal neural network (MSNN) model. In this method, images are classified into five categories, i.e., “Car", “Building", “Mountain", “Farm" and “Coast". This classification is performed without any segmentation processes. To verify the learning capabilities of the proposed method, we compare our MSNN model with the traditional Sigmoidal Neural Network (SNN) model. Results of comparison have shown that the MSNN model performs better than the traditional SNN model in the context of training run time and classification rate. Both color moments and multi-level wavelets decomposition technique are used to extract features from images. The proposed method has been tested on a variety of real and synthetic images.

Keywords: Image classification, multi-level neural networks, feature extraction, wavelets decomposition.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1602

Authors: Yuanyuan Chai, Limin Jia, Zundong Zhang

Abstract:

Hybrid algorithm is the hot issue in Computational Intelligence (CI) study. From in-depth discussion on Simulation Mechanism Based (SMB) classification method and composite patterns, this paper presents the Mamdani model based Adaptive Neural Fuzzy Inference System (M-ANFIS) and weight updating formula in consideration with qualitative representation of inference consequent parts in fuzzy neural networks. M-ANFIS model adopts Mamdani fuzzy inference system which has advantages in consequent part. Experiment results of applying M-ANFIS to evaluate traffic Level of service show that M-ANFIS, as a new hybrid algorithm in computational intelligence, has great advantages in non-linear modeling, membership functions in consequent parts, scale of training data and amount of adjusted parameters.

Keywords: Fuzzy neural networks, Mamdani fuzzy inference, M-ANFIS

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 5191

Authors: Meng Hu, Lili Wang

Abstract:

In this paper, we investigate dynamics of 2n almost periodic attractors for Cohen-Grossberg neural networks (CGNNs) with variable and distribute time delays. By imposing some new assumptions on activation functions and system parameters, we split invariant basin of CGNNs into 2n compact convex subsets. Then the existence of 2n almost periodic solutions lying in compact convex subsets is attained due to employment of the theory of exponential dichotomy and Schauder-s fixed point theorem. Meanwhile, we derive some new criteria for the networks to converge toward these 2n almost periodic solutions and exponential attracting domains are also given correspondingly.

Keywords: CGNNs, almost periodic solution, invariant basins, attracting domains.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1331

Authors: Hidehiko Okada

Abstract:

The author previously proposed an extension of differential evolution. The proposed method extends the processes of DE to handle interval numbers as genotype values so that DE can be applied to interval-valued optimization problems. The interval DE can employ either of two interval models, the lower and upper model or the center and width model, for specifying genotype values. Ability of the interval DE in searching for solutions may depend on the model. In this paper, the author compares the two models to investigate which model contributes better for the interval DE to find better solutions. Application of the interval DE is evolutionary training of interval-valued neural networks. A result of preliminary study indicates that the CW model is better than the LU model: the interval DE with the CW model could evolve better neural networks. 

Keywords: Evolutionary algorithms, differential evolution, neural network, neuroevolution, interval arithmetic.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1612

Authors: Vishnu Pratap Singh Kirar

Abstract:

The main aim of a communication system is to achieve maximum performance. In Cognitive Radio any user or transceiver has ability to sense best suitable channel, while channel is not in use. It means an unlicensed user can share the spectrum of a licensed user without any interference. Though, the spectrum sensing consumes a large amount of energy and it can reduce by applying various artificial intelligent methods for determining proper spectrum holes. It also increases the efficiency of Cognitive Radio Network (CRN). In this survey paper we discuss the use of different learning models and implementation of Artificial Neural Network (ANN) to increase the learning and decision making capacity of CRN without affecting bandwidth, cost and signal rate.

Keywords: Artificial Neural Network, Cognitive Radio, Cognitive Radio Networks, Back Propagation, Spectrum Sensing.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 4050

Authors: Mengzhuo Luo, Shouming Zhong

Abstract:

This paper investigates the problem of exponential stability for a class of uncertain discrete-time stochastic neural network with time-varying delays. By constructing a suitable Lyapunov-Krasovskii functional, combining the stochastic stability theory, the free-weighting matrix method, a delay-dependent exponential stability criteria is obtained in term of LMIs. Compared with some previous results, the new conditions obtain in this paper are less conservative. Finally, two numerical examples are exploited to show the usefulness of the results derived.

Keywords: Delay-dependent stability, Neural networks, Time varying delay, Linear matrix inequality (LMI).

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1888

Authors: Meng Hu, Lili Wang

Abstract:

In this paper, based on linear matrix inequality (LMI), by using Lyapunov functional theory, the exponential stability criterion is obtained for a class of uncertain Takagi-Sugeno fuzzy Hopfield neural networks (TSFHNNs) with time delays. Here we choose a generalized Lyapunov functional and introduce a parameterized model transformation with free weighting matrices to it, these techniques lead to generalized and less conservative stability condition that guarantee the wide stability region. Finally, an example is given to illustrate our results by using MATLAB LMI toolbox.

Keywords: Hopfield neural network, linear matrix inequality, exponential stability, time delay, T-S fuzzy model.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1462

Authors: Daniel Rivero, Julián Dorado, Juan R. Rabuñal, Alejandro Pazos, Javier Pereira

Abstract:

The development of Artificial Neural Networks (ANNs) is usually a slow process in which the human expert has to test several architectures until he finds the one that achieves best results to solve a certain problem. This work presents a new technique that uses Genetic Programming (GP) for automatically generating ANNs. To do this, the GP algorithm had to be changed in order to work with graph structures, so ANNs can be developed. This technique also allows the obtaining of simplified networks that solve the problem with a small group of neurons. In order to measure the performance of the system and to compare the results with other ANN development methods by means of Evolutionary Computation (EC) techniques, several tests were performed with problems based on some of the most used test databases. The results of those comparisons show that the system achieves good results comparable with the already existing techniques and, in most of the cases, they worked better than those techniques.

Keywords: Artificial Neural Networks, Evolutionary Computation, Genetic Programming.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1416

Authors: Meng Hu, Lili Wang

Abstract:

In this paper, based on the estimation of the Cauchy matrix of linear impulsive differential equations, by using Banach fixed point theorem and Gronwall-Bellman-s inequality, some sufficient conditions are obtained for the existence and exponential stability of almost periodic solution for Cohen-Grossberg shunting inhibitory cellular neural networks (SICNNs) with continuously distributed delays and impulses. An example is given to illustrate the main results.

Keywords: Almost periodic solution, exponential stability, neural networks, impulses.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 14287

Authors: Truong Quang Dang Khoa, Masahiro Nakagawa

Abstract:

Brain Computer Interface (BCI) has been recently increased in research. Functional Near Infrared Spectroscope (fNIRs) is one the latest technologies which utilize light in the near-infrared range to determine brain activities. Because near infrared technology allows design of safe, portable, wearable, non-invasive and wireless qualities monitoring systems, fNIRs monitoring of brain hemodynamics can be value in helping to understand brain tasks. In this paper, we present results of fNIRs signal analysis indicating that there exist distinct patterns of hemodynamic responses which recognize brain tasks toward developing a BCI. We applied two different mathematics tools separately, Wavelets analysis for preprocessing as signal filters and feature extractions and Neural networks for cognition brain tasks as a classification module. We also discuss and compare with other methods while our proposals perform better with an average accuracy of 99.9% for classification.

Keywords: functional near infrared spectroscope (fNIRs), braincomputer interface (BCI), wavelets, neural networks, brain activity, neuroimaging.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1976

Authors: Isao Taguchi, Yasuo Sugai

Abstract:

This paper proposes an efficient learning method for the layered neural networks based on the selection of training data and input characteristics of an output layer unit. Comparing to recent neural networks; pulse neural networks, quantum neuro computation, etc, the multilayer network is widely used due to its simple structure. When learning objects are complicated, the problems, such as unsuccessful learning or a significant time required in learning, remain unsolved. Focusing on the input data during the learning stage, we undertook an experiment to identify the data that makes large errors and interferes with the learning process. Our method devides the learning process into several stages. In general, input characteristics to an output layer unit show oscillation during learning process for complicated problems. The multi-stage learning method proposes by the authors for the function approximation problems of classifying learning data in a phased manner, focusing on their learnabilities prior to learning in the multi layered neural network, and demonstrates validity of the multi-stage learning method. Specifically, this paper verifies by computer experiments that both of learning accuracy and learning time are improved of the BP method as a learning rule of the multi-stage learning method. In learning, oscillatory phenomena of a learning curve serve an important role in learning performance. The authors also discuss the occurrence mechanisms of oscillatory phenomena in learning. Furthermore, the authors discuss the reasons that errors of some data remain large value even after learning, observing behaviors during learning.

Keywords: data selection, function approximation problem, multistage leaning, neural network, voluntary oscillation.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1386

Authors: Salvatore Marra, Francesco C. Morabito

Abstract:

In this paper we present an efficient approach for the prediction of two sunspot-related time series, namely the Yearly Sunspot Number and the IR5 Index, that are commonly used for monitoring solar activity. The method is based on exploiting partially recurrent Elman networks and it can be divided into three main steps: the first one consists in a “de-rectification" of the time series under study in order to obtain a new time series whose appearance, similar to a sum of sinusoids, can be modelled by our neural networks much better than the original dataset. After that, we normalize the derectified data so that they have zero mean and unity standard deviation and, finally, train an Elman network with only one input, a recurrent hidden layer and one output using a back-propagation algorithm with variable learning rate and momentum. The achieved results have shown the efficiency of this approach that, although very simple, can perform better than most of the existing solar activity forecasting methods.

Keywords: Elman neural networks, sunspot, solar activity, time series prediction.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1810

Authors: F. Allag, S. Bouharati, M. Belmahdi, R. Zegadi

Abstract:

The rapid expansion of deserts in recent decades as a result of human actions combined with climatic changes has highlighted the necessity to understand biological processes in arid environments. Whereas physical processes and the biology of flora and fauna have been relatively well studied in marginally used arid areas, knowledge of desert soil micro-organisms remains fragmentary. The objective of this study is to conduct a diversity analysis of bacterial communities in unvegetated arid soils. Several biological phenomena in hot deserts related to microbial populations and the potential use of micro-organisms for restoring hot desert environments. Dry land ecosystems have a highly heterogeneous distribution of resources, with greater nutrient concentrations and microbial densities occurring in vegetated than in bare soils. In this work, we found it useful to use techniques of artificial intelligence in their treatment especially artificial neural networks (ANN). The use of the ANN model, demonstrate his capability for addressing the complex problems of uncertainty data.

Keywords: Desert soil, Climatic changes, Bacteria, Vegetation, Artificial neural networks.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1854

Authors: O. Yavuz, L. Ozyilmaz

Abstract:

HIV-1 genome is highly heterogeneous. Due to this variation, features of HIV-I genome is in a wide range. For this reason, the ability to infection of the virus changes depending on different chemokine receptors. From this point of view, R5 HIV viruses use CCR5 coreceptor while X4 viruses use CXCR5 and R5X4 viruses can utilize both coreceptors. Recently, in Bioinformatics, R5X4 viruses have been studied to classify by using the experiments on HIV-1 genome. In this study, R5X4 type of HIV viruses were classified using Auto Regressive (AR) model through Artificial Neural Networks (ANNs). The statistical data of R5X4, R5 and X4 viruses was analyzed by using signal processing methods and ANNs. Accessible residues of these virus sequences were obtained and modeled by AR model since the dimension of residues is large and different from each other. Finally the pre-processed data was used to evolve various ANN structures for determining R5X4 viruses. Furthermore ROC analysis was applied to ANNs to show their real performances. The results indicate that R5X4 viruses successfully classified with high sensitivity and specificity values training and testing ROC analysis for RBF, which gives the best performance among ANN structures.

Keywords: Auto-Regressive Model, HIV, Neural Networks, ROC Analysis.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1138

Authors: S. Anna Durai, E. Anna Saro

Abstract:

In general the images used for compression are of different types like dark image, high intensity image etc. When these images are compressed using Counter Propagation Neural Network, it takes longer time to converge. The reason for this is that the given image may contain a number of distinct gray levels with narrow difference with their neighborhood pixels. If the gray levels of the pixels in an image and their neighbors are mapped in such a way that the difference in the gray levels of the neighbor with the pixel is minimum, then compression ratio as well as the convergence of the network can be improved. To achieve this, a Cumulative Distribution Function is estimated for the image and it is used to map the image pixels. When the mapped image pixels are used the Counter Propagation Neural Network yield high compression ratio as well as it converges quickly.

Keywords: Correlation, Counter Propagation Neural Networks, Cummulative Distribution Function, Image compression.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1623

Authors: G. R. Rameshkumar, B. V. A Rao, K. P. Ramachandran

Abstract:

Misalignment and unbalance are the major concerns in rotating machinery. When the power supply to any rotating system is cutoff, the system begins to lose the momentum gained during sustained operation and finally comes to rest. The exact time period from when the power is cutoff until the rotor comes to rest is called Coast Down Time. The CDTs for different shaft cutoff speeds were recorded at various misalignment and unbalance conditions. The CDT reduction percentages were calculated for each fault and there is a specific correlation between the CDT reduction percentage and the severity of the fault. In this paper, radial basis network, a new generation of artificial neural networks, has been successfully incorporated for the prediction of CDT for misalignment and unbalance conditions. Radial basis network has been found to be successful in the prediction of CDT for mechanical faults in rotating machinery.

Keywords: Coast Down Time, Misalignment, Unbalance, Artificial Neural Networks, Radial Basis Network.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2933

Authors: Wei Zhang

Abstract:

With the rapid development of deep learning, neural network and deep learning algorithms play a significant role in various practical applications. Due to the high accuracy and good performance, Convolutional Neural Networks (CNNs) especially have become a research hot spot in the past few years. However, the size of the networks becomes increasingly large scale due to the demands of the practical applications, which poses a significant challenge to construct a high-performance implementation of deep learning neural networks. Meanwhile, many of these application scenarios also have strict requirements on the performance and low-power consumption of hardware devices. Therefore, it is particularly critical to choose a moderate computing platform for hardware acceleration of CNNs. This article aimed to survey the recent advance in Field Programmable Gate Array (FPGA)-based acceleration of CNNs. Various designs and implementations of the accelerator based on FPGA under different devices and network models are overviewed, and the versions of Graphic Processing Units (GPUs), Application Specific Integrated Circuits (ASICs) and Digital Signal Processors (DSPs) are compared to present our own critical analysis and comments. Finally, we give a discussion on different perspectives of these acceleration and optimization methods on FPGA platforms to further explore the opportunities and challenges for future research. More helpfully, we give a prospect for future development of the FPGA-based accelerator.

Keywords: Deep learning, field programmable gate array, FPGA, hardware acceleration, convolutional neural networks, CNN.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 814

Authors: Ridha Ghayoula, Mbarek Traii, Ali Gharsallah

Abstract:

In this paper, we intend to study the synthesis of the multibeam arrays. The synthesis implementation-s method for this type of arrays permits to approach the appropriated radiance-s diagram. The used approach is based on neural network that are capable to model the multibeam arrays, consider predetermined general criteria-s, and finally it permits to predict the appropriated diagram from the neural model. Our main contribution in this paper is the extension of a synthesis model of these multibeam arrays.

Keywords: Multibeam, modelling, neural networks, synthesis, antennas.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1181

Authors: Yasser F. Hassan

Abstract:

Deep learning structure is a branch of machine learning science and greet achievement in research and applications. Cellular neural networks are regarded as array of nonlinear analog processors called cells connected in a way allowing parallel computations. The paper discusses how to use deep learning structure for representing neural cellular automata model. The proposed learning technique in cellular automata model will be examined from structure of deep learning. A deep automata neural cellular system modifies each neuron based on the behavior of the individual and its decision as a result of multi-level deep structure learning. The paper will present the architecture of the model and the results of simulation of approach are given. Results from the implementation enrich deep neural cellular automata system and shed a light on concept formulation of the model and the learning in it.

Keywords: Cellular automata, neural cellular automata, deep learning, classification.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 814

Authors: Biao Qin, Jin Huang

Abstract:

In this paper, together with some improved Lyapunov-Krasovskii functional and effective mathematical techniques, several sufficient conditions are derived to guarantee the error system is globally asymptotically stable with H∞ performance, in which both the time-delay and its time variation can be fully considered. In order to get less conservative results of the state estimation condition, zero equalities and reciprocally convex approach are employed. The estimator gain matrix can be obtained in terms of the solution to linear matrix inequalities. A numerical example is provided to illustrate the usefulness and effectiveness of the obtained results.

Keywords: H∞ performance, Neural networks, State estimation.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1400

Authors: A. Pashayev, R. Sadiqov, C. Ardil, F. Ildiz , H. Karabork

Abstract:

The new technology of fuzzy neural networks for identification of parameters for mathematical models of geofields is proposed and checked. The effectiveness of that soft computing technology is demonstrated, especially in the early stage of modeling, when the information is uncertain and limited.

Keywords: Identification, interpolation methods, neuro-fuzzy networks, geofield.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1286

Authors: Diego Ordonez, Carlos Dafonte, Minia Manteiga, Bernardino Arcayy

Abstract:

This work presents a neural network model for the clustering analysis of data based on Self Organizing Maps (SOM). The model evolves during the training stage towards a hierarchical structure according to the input requirements. The hierarchical structure symbolizes a specialization tool that provides refinements of the classification process. The structure behaves like a single map with different resolutions depending on the region to analyze. The benefits and performance of the algorithm are discussed in application to the Iris dataset, a classical example for pattern recognition.

Keywords: Neural networks, Self-organizing feature maps, Hierarchicalsystems, Pattern clustering methods.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1896

Authors: Shivakumar, G. S. Vijay, P. Srinivas Pai, B. R. Shrinivasa Rao

Abstract:

In the present study, RBF neural networks were used for predicting the performance and emission parameters of a biodiesel engine. Engine experiments were carried out in a 4 stroke diesel engine using blends of diesel and Honge methyl ester as the fuel. Performance parameters like BTE, BSEC, Tex and emissions from the engine were measured. These experimental results were used for ANN modeling. RBF center initialization was done by random selection and by using Clustered techniques. Network was trained by using fixed and varying widths for the RBF units. It was observed that RBF results were having a good agreement with the experimental results. Networks trained by using clustering technique gave better results than using random selection of centers in terms of reduced MRE and increased prediction accuracy. The average MRE for the performance parameters was 3.25% with the prediction accuracy of 98% and for emissions it was 10.4% with a prediction accuracy of 80%.

Keywords: Radial Basis Function networks, emissions, Performance parameters, Fuzzy c means.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 1684

Authors: Nnamdi I. Nwulu, Shola G. Oroja

Abstract:

It has become crucial over the years for nations to improve their credit scoring methods and techniques in light of the increasing volatility of the global economy. Statistical methods or tools have been the favoured means for this; however artificial intelligence or soft computing based techniques are becoming increasingly preferred due to their proficient and precise nature and relative simplicity. This work presents a comparison between Support Vector Machines and Artificial Neural Networks two popular soft computing models when applied to credit scoring. Amidst the different criteria-s that can be used for comparisons; accuracy, computational complexity and processing times are the selected criteria used to evaluate both models. Furthermore the German credit scoring dataset which is a real world dataset is used to train and test both developed models. Experimental results obtained from our study suggest that although both soft computing models could be used with a high degree of accuracy, Artificial Neural Networks deliver better results than Support Vector Machines.

Keywords: Artificial Neural Networks, Credit Scoring, SoftComputing Models, Support Vector Machines.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2089

Authors: Abderrahmane Amrouche, Jean Michel Rouvaen

Abstract:

In this paper we present an efficient system for independent speaker speech recognition based on neural network approach. The proposed architecture comprises two phases: a preprocessing phase which consists in segmental normalization and features extraction and a classification phase which uses neural networks based on nonparametric density estimation namely the general regression neural network (GRNN). The relative performances of the proposed model are compared to the similar recognition systems based on the Multilayer Perceptron (MLP), the Recurrent Neural Network (RNN) and the well known Discrete Hidden Markov Model (HMM-VQ) that we have achieved also. Experimental results obtained with Arabic digits have shown that the use of nonparametric density estimation with an appropriate smoothing factor (spread) improves the generalization power of the neural network. The word error rate (WER) is reduced significantly over the baseline HMM method. GRNN computation is a successful alternative to the other neural network and DHMM.

Keywords: Speech Recognition, General Regression NeuralNetwork, Hidden Markov Model, Recurrent Neural Network, ArabicDigits.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 2134

Authors: Zhifeng Kong

Abstract:

Over-parameterized neural networks have attracted a great deal of attention in recent deep learning theory research, as they challenge the classic perspective of over-fitting when the model has excessive parameters and have gained empirical success in various settings. While a number of theoretical works have been presented to demystify properties of such models, the convergence properties of such models are still far from being thoroughly understood. In this work, we study the convergence properties of training two-hidden-layer partially over-parameterized fully connected networks with the Rectified Linear Unit activation via gradient descent. To our knowledge, this is the first theoretical work to understand convergence properties of deep over-parameterized networks without the equally-wide-hidden-layer assumption and other unrealistic assumptions. We provide a probabilistic lower bound of the widths of hidden layers and proved linear convergence rate of gradient descent. We also conducted experiments on synthetic and real-world datasets to validate our theory.

Keywords: Over-parameterization, Rectified Linear Units (ReLU), convergence, gradient descent, neural networks.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 811

Authors: Apoorva Vinod, Archana Mathur, Snehanshu Saha

Abstract:

Though there exists a plethora of Activation Functions (AFs) used in single and multiple hidden layer Neural Networks (NN), their behavior always raised curiosity, whether used in combination or singly. The popular AFs – Sigmoid, ReLU, and Tanh – have performed prominently well for shallow and deep architectures. Most of the time, AFs are used singly in multi-layered NN, and, to the best of our knowledge, their performance is never studied and analyzed deeply when used in combination. In this manuscript, we experiment on multi-layered NN architecture (both on shallow and deep architectures; Convolutional NN and VGG16) and investigate how well the network responds to using two different AFs (Sigmoid-Tanh, Tanh-ReLU, ReLU-Sigmoid) used alternately against a traditional, single (Sigmoid-Sigmoid, Tanh-Tanh, ReLU-ReLU) combination. Our results show that on using two different AFs, the network achieves better accuracy, substantially lower loss, and faster convergence on 4 computer vision (CV) and 15 Non-CV (NCV) datasets. When using different AFs, not only was the accuracy greater by 6-7%, but we also accomplished convergence twice as fast. We present a case study to investigate the probability of networks suffering vanishing and exploding gradients when using two different AFs. Additionally, we theoretically showed that a composition of two or more AFs satisfies Universal Approximation Theorem (UAT).

Keywords: Activation Function, Universal Approximation function, Neural Networks, convergence.

Procedia APA BibTeX Chicago EndNote Harvard JSON MLA RIS XML ISO 690 PDF Downloads 87